Project

General

Profile

Actions

Feature #7792

closed

Make symbols and strings the same thing

Added by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago. Updated over 6 years ago.

Status:
Rejected
Target version:
[ruby-core:84474]

Description

Recently I had to replace several of my symbols to plain strings in my project. Here is what happened:

I generated some results with some class that would add items to an array like this:

results << {id: 1, name: 'abc'}

Then I would store such results in cache using Redis, encoded as a JSON string. But then when I restore the data from cache the hash will be {'id' => 1, 'name' => 'abc'}.

This wasn't a problem until recently because I never used the results directly in the same request before and would always use the value stored on Redis and parsed by JSON.

But recently I had to use the values directly in a view. But then I had a problem because I would have to use symbols in the results for the first time and strings the next times when the result was available on cache. I really don't want to care about memory management in Ruby if possible and symbols forces me to abandon the new sexy hash syntax many times. Now I have to write

results << {'id' => 1, 'name' => 'abc}

when I'd prefer to write

results << {id: 1, name: 'abc}

This is not the first time I had a bad user experience due to symbols being different from strings. And I'm not the only one or ActiveSupport Hash#with_indifferent_access wouldn't be so popular and Rails wouldn't use it all the time internally.

It is really bad when you have to constantly check how you store your keys in your hashes. Am I using symbols or strings as keys? If you use the wrong type on plain hashes you can find a bad time debugging your code. Or you could just use Hash#with_indifferent_access everywhere, thus reducing performance (I guess) and it is pretty inconvenient anyway.

Or if you're comparing the keys of your hash in some "each" closure you have to worry about it being a symbol or a string too.

Ruby is told to be programmers' friendly and it usually is. But symbols are certainly a big exception.


Related issues 1 (0 open1 closed)

Related to Ruby master - Feature #5964: Make Symbols an Alternate Syntax for StringsRejectedmatz (Yukihiro Matsumoto)Actions

Updated by trans (Thomas Sawyer) about 11 years ago

I'm not sure that's even possible. If String#hash produced the same number as Symbol#hash, then that would do the trick, but that probably lead to some unforeseen breakages.

I do know one related thing, though. I don't like having to do:

if (String === x || Symbol === x)
  x.to_s

When all I really want to do is x.to_str, but handle Symbols too.

Oh, btw, I've suggested that Hash add a convert_key procedure which could be used to normalize keys. Very useful, but would obviously mean a bit of speed hit.

Updated by luislavena (Luis Lavena) about 11 years ago

rosenfeld: see #5964 for similar discussion.

Updated by shyouhei (Shyouhei Urabe) about 11 years ago

Mmm, it sounds too big to me.

@rosenfeld (Rodrigo Rosenfeld Rosas) I know your situation. But is it really a right solution for you? How about a hash with indifferent access? Or how about changing {foo:1} to be { 'foo' => 1 }, not { :foo => 1 } ? It seems your frustration can be relaxed in several ways. Letting symbols be strings is not the only one and (seems to be) suboptiomal.

Updated by yorickpeterse (Yorick Peterse) about 11 years ago

Symbols and Strings both have different use cases and that's actually a
good thing. If you want to be able to use both Strings and Symbols as
your Hash keys you can use something like Hashie:
https://github.com/intridea/hashie

Yorick

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Although I'd really prefer that symbols and strings were the same there is an alternative that would satisfy me as well:

Make Hash behave as HashWithIndifferentAccess and create a new class StrictHash to keep the behavior of the existent Hash class. That way this would work:

a = {a: 1}; a[:a] == a['a']

Also any stdlib libraries, such as JSON, should use Hash instead of StrictHash on parse.

That way it wouldn't really matter if {foo: 1} maps to {:foo => 1} or {'foo' => 1}. By the way I'm still curious to know if we'll ever be able to use string interpolation in the hash syntax, like {"#{'foo'}": 1}.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

@yorickpeterse (Yorick Peterse), your suggestion wouldn't work for my case. The hash is created by JSON.parse where I don't control the hash creation. And I don't like the idea of monkey patching core classes either. Specially in fundamental classes like Hash. If I patch it to make it behave like HashWithIndifferentAccess it could break many libraries/frameworks in unexpected ways.

Updated by yorickpeterse (Yorick Peterse) about 11 years ago

You don't need to hijack any code for it, you'd just use it as
following:

  require 'hashie'

  parsed = JSON.parse('{"name": "Ruby"}')
  hash   = Hashie::Mash.new(parsed)

  hash.name    # => "Ruby"
  hash['name'] # => "Ruby"
  hash[:name]  # => "Ruby"

We use Hashie in various production applications and it works quite well
for us.

Yorick

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 06-02-2013 12:36, Yorick Peterse escreveu:

You don't need to hijack any code for it, you'd just use it as
following:

require 'hashie'

parsed = JSON.parse('{"name": "Ruby"}')
hash   = Hashie::Mash.new(parsed)

hash.name    # => "Ruby"
hash['name'] # => "Ruby"
hash[:name]  # => "Ruby"

We use Hashie in various production applications and it works quite well
for us.

So you get a performance hit because you don't want to worry about
symbols while symbols are meant to give you better performance, right?
How ironic is that?

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

I'd just like to highlight what performance impact we may be referring to:

https://gist.github.com/rosenfeld/4723061

I'll copy it here:

require 'benchmark'

hashes = []
1_000_000.times { hashes << { some_key_name: 1, 'some_key_name' => 2 }}
Benchmark.bmbm do |x|
x.report { hashes.map{|h| h[:some_key_name]} }
x.report { hashes.map{|h| h['some_key_name']} }
end

Result:

ruby symbols-performance.rb
Rehearsal ------------------------------------
0.540000 0.010000 0.550000 ( 0.543421)
0.680000 0.020000 0.700000 ( 0.705931)
--------------------------- total: 1.250000sec

   user     system      total        real

0.240000 0.000000 0.240000 ( 0.244554)
0.380000 0.000000 0.380000 ( 0.391517)

Does this small hit in performance worth all the hassle that it is to have to differentiate symbols from strings?

Updated by yorickpeterse (Yorick Peterse) about 11 years ago

I don't think I'm following you, can you explain what's supposedly
ironic about it? Using Hashie only "slows" things down based on whether
you use Symbols, Strings or object attributes. Unless you use it all
over the place the performance impact is small.

I personally don't fully agree with what Hashie does because I believe
people should be competent enough to realize that when they take in
external data it's going to be String instances (for keys that is).

Having said that, I think fundamentally changing the way Ruby works when
it comes to handling Strings and Symbols because developers can't be
bothered fixing the root cause of the problem is flawed. If you're
worried about a ddos stop converting everything to Symbols. If you're
worried about not remember what key type to use, use a custom object or
document it so that people can easily know.

While Ruby is all about making the lifes easier I really don't want it
to become a language that spoon feeds programmers because they're too
lazy to type 1 extra character or convert the output manually. Or
better: use a custom object as mention above.

The benchmark you posted is flawed because it does much, much more than
benchmarking the time required to create a new Symbol or String
instance. Lets take a look at the most basic benchmark of these two data
types:

  require 'benchmark'

  amount = 50000000

  Benchmark.bmbm(40) do |run|
    run.report 'Symbols' do
      amount.times do
        :foobar
      end
    end

    run.report 'Strings' do
      amount.times do
        'foobar'
      end
    end
  end

On the laptop I'm currently using this results in the following output:

  Rehearsal 

  Symbols                                    2.310000   0.000000 

2.310000 ( 2.311325)
Strings 5.710000 0.000000
5.710000 ( 5.725365)
-------------------------------------------------------------------
total: 8.020000sec

                                                 user     system 

total real
Symbols 2.670000 0.000000
2.670000 ( 2.680489)
Strings 6.560000 0.010000
6.570000 ( 6.584651)

This shows that the use of Strings is roughly 2,5 times slower than
Symbols. Now execution time isn't the biggest concern in this case, it's
memory usage. For this I used the following basic benchmark:

  def get_memory
    return `ps -o rss= #{Process.pid}`.strip.to_f
  end

  def benchmark_memory
    before = get_memory

    yield

    return get_memory - before
  end

  amount = 50000000

  puts "Start memory: #{get_memory} KB"

  symbols = benchmark_memory do
    amount.times do
      :foobar
    end
  end

  strings = benchmark_memory do
    amount.times do
      'foobar'
    end
  end

  puts "Symbols used #{symbols} KB"
  puts "Strings used #{strings} KB"

This results in the following:

  Start memory: 4876.0 KB
  Symbols used 0.0 KB
  Strings used 112.0 KB

Now I wouldn't be too surprised if there's some optimization going on
because I'm re-creating the same values over and over again but it
already shows a big difference between the two.

To cut a long story short: I can understand what you're trying to get
at, both with the two data types being merged and the ddos issue.
However, I feel neither of these issues are an issue directly related to
Ruby itself. If Ruby were to automatically convert things to Symbols for
you then yes, but in this case frameworks such as Rails are the cause of
the problem. Merging the two datatypes would most likely make such a
huge different usage/code wise that it would probably be something for
Ruby 5.0 (in other words, not in the near future).

Yorick

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 06-02-2013 13:25, Yorick Peterse escreveu:

I don't think I'm following you, can you explain what's supposedly
ironic about it? Using Hashie only "slows" things down based on whether
you use Symbols, Strings or object attributes. Unless you use it all
over the place the performance impact is small.

What I'm trying to say is that the main reason why symbols exist in Ruby
in the first place is performance from what I've been told.

But then people don't want to worry if hashes are indexed by strings or
symbols so they end up using some kind of HashWithIndifferentAccess or
similar techniques. But since the normal Hash class doesn't behave this
way you have to loop through all hashes in an object returned by
JSON.parse to make them behave as HashWithIndifferentAccess, which is
has a huge performance hit when compared to the small gains symbols
could add.

I personally don't fully agree with what Hashie does because I believe
people should be competent enough to realize that when they take in
external data it's going to be String instances (for keys that is).

It is not a matter of being competent or not. You can't know in advance
if a hash returned by some external code is indexed by string or
symbols. You have to test by yourself or check the documentation. Or you
could just use a HashWithIndifferentAccess class and stop worrying about
it. This has a big impact on coding speed and software maintenance,
which is the big problem in my opinion.

Having said that, I think fundamentally changing the way Ruby works when
it comes to handling Strings and Symbols because developers can't be
bothered fixing the root cause of the problem is flawed.

People reading some Ruby book will notice that it is not particularly
designed with performance in mind but it is designed mostly towards
programmer's happiness. If that is the case, then worrying about
bothered programmers makes sense to a language like Ruby in my opinion.

If you're worried about a ddos

DDoS is a separate beast that can't be easily prevented no matter what
language/framework you use. I'm just talking about DoS exploiting
through memory exhaustion due to symbols not being collected. Anyway,
this is a separate issue from this one and would be better discussed in
that separate thread.

stop converting everything to Symbols.

I'm not converting anything to symbols. Did you read the feature
description use case? I'm just creating regular hashes using the new
sexy hash syntax which happens to create symbols instead of strings.
Then when I serialize my object to JSON for storing on Redis for caching
purpose I'll get a hash indexed by strings instead of symbols. That
means that when I'm accessing my hash I have to be worried if the hash
has been just generated or if it was loaded from Redis to decide if I
should use strings or symbols to get the hash values. There are many
more similar situations where this difference between symbols and
strings will cause confusion. And I don't see much benefits in keeping
them separate things either.

If you're worried about not remember what key type to use, use a
custom object or
document it so that people can easily know.

This isn't possible when you're serializing/deserializing using some
library like JSON or any other. You don't control how hashes are created
by such libraries.

While Ruby is all about making the lifes easier I really don't want it
to become a language that spoon feeds programmers because they're too
lazy to type 1 extra character or convert the output manually. Or
better: use a custom object as mention above.

Again, see the ticket description first before assuming things.

The benchmark you posted is flawed because it does much, much more than
benchmarking the time required to create a new Symbol or String
instance. Lets take a look at the most basic benchmark of these two data
types:

require 'benchmark'

amount = 50000000

Benchmark.bmbm(40) do |run|
  run.report 'Symbols' do
    amount.times do
      :foobar
    end
  end

  run.report 'Strings' do
    amount.times do
      'foobar'
    end
  end
end

On the laptop I'm currently using this results in the following output:

Rehearsal 

Symbols                                    2.310000   0.000000 

2.310000 ( 2.311325)
Strings 5.710000 0.000000
5.710000 ( 5.725365)


total: 8.020000sec

                                               user     system 

total real
Symbols 2.670000 0.000000
2.670000 ( 2.680489)
Strings 6.560000 0.010000
6.570000 ( 6.584651)

This shows that the use of Strings is roughly 2,5 times slower than
Symbols. Now execution time isn't the biggest concern in this case, it's
memory usage.

Exactly, no real-world software would consist mostly of creating
strings/symbols. Even in a simplistic context like my example, it is
hard to notice any impact on the overall code caused by string
allocation taking more time than symbols.When we get more complete code
we'll notice that it really doesn't make any difference if we're using
symbols or strings all over our code...

Also, any improvements on threading and parallelizing support are likely
to yield much bigger performance boots than any micro-optimization with
symbols instead of strings.

For this I used the following basic benchmark:

def get_memory
  return `ps -o rss= #{Process.pid}`.strip.to_f
end

def benchmark_memory
  before = get_memory

  yield

  return get_memory - before
end

amount = 50000000

puts "Start memory: #{get_memory} KB"

symbols = benchmark_memory do
  amount.times do
    :foobar
  end
end

strings = benchmark_memory do
  amount.times do
    'foobar'
  end
end

puts "Symbols used #{symbols} KB"
puts "Strings used #{strings} KB"

This results in the following:

Start memory: 4876.0 KB
Symbols used 0.0 KB
Strings used 112.0 KB

Now I wouldn't be too surprised if there's some optimization going on
because I'm re-creating the same values over and over again but it
already shows a big difference between the two.

112KB isn't certainly a big difference in my opinion unless you're
designing some embedded application. I've worked with embedded devices
in the past and although I see some attempts to make a lighter Ruby
subset (like mRuby) for such use-case I'd certainly use C or C++ for my
embedded apps these days. Did you know that Java initially was supposed
to be used by embedded devices from what I've been told? Then it tried
to convince people to use it to create multi-platform desktop apps.
After that its initial footprint was so big that it wasn't a good idea
to try it on embedded devices for most cases. Then they tried to make it
work in browsers through applets. Now it seems people want to use Java
mostly for web servers (HTTP and other protocols). The result was a big
mess in my opinion. I don't think Ruby (the full specification) should
be concerned about embedded devices. C is already a good fit for devices
with small memory constraints. When you consider using Ruby it is likely
that you have more CPU and memory resources than a typical small device
would have, so 112KB wouldn't make much difference.

And for embedded devices, it is also recommended that they run some RTOS
instead of plain Linux. If they want to keep with Linux, an option would
be to patch it with Xenomai patch for instance. But in that case, any
real-time task would be implemented in C, not in Ruby or any other
language subjected to garbage collected, like Java. So, if we keep the
focus on applications running on normal computers, 112KB won't really
make any difference, don't you agree?

To cut a long story short: I can understand what you're trying to get
at, both with the two data types being merged and the ddos issue.
However, I feel neither of these issues are an issue directly related to
Ruby itself. If Ruby were to automatically convert things to Symbols for
you then yes, but in this case frameworks such as Rails are the cause of
the problem.

Rails is not related at all to the use case I pointed out in this ticket
description. It happens with regular Ruby classes (JSON, Hash) and with
the "redis" gem that is independent from Rails.

Merging the two datatypes would most likely make such a
huge different usage/code wise that it would probably be something for
Ruby 5.0 (in other words, not in the near future).

Ruby 3.0 won't happen in a near future. Next Major means Ruby 3.0 if I
understand it correctly.

Updated by yorickpeterse (Yorick Peterse) about 11 years ago

What I'm trying to say is that the main reason why symbols exist in
Ruby in the first place is performance from what I've been told.

Correct, and your proposed changes would completely nullify those
performance benefits (see below).

People reading some Ruby book will notice that it is not particularly
designed with performance in mind but it is designed mostly towards
programmer's happiness. If that is the case, then worrying about
bothered programmers makes sense to a language like Ruby in my
opinion.

So basically what you're saying is "Ruby is written for happiness and
not performance, lets make it even more slow!". I'd rather see a world
where Ruby is both fast (enough) and easy to use instead of it being
easy to use and slower than a sloth.

Regarding the benchmarking information, you're missing a crucial aspect.
While the numbers in the specific examples I gave both clearly show that
the use of Strings is substantially slower. Yes, it's "only" 112 kb but
the difference will keep growing and growing until you hit your memory
limit.

This is exactly one of the reasons Symbols exist: to make it easier and
faster to use commonly re-used Strings. The best example of this are
Hash keys.

This isn't possible when you're serializing/deserializing using some
library like JSON or any other. You don't control how hashes are
created by such libraries.

Of course it is. Marshal allows you to store arbitrary Ruby objects
(with the exception of a few such as Proc instances), in other cases you
can re-create your objects based on the supplied Hash.

If you do not like using raw Hashes the solution in my opinion is not to
more or less re-write Ruby (and break everything that exists in the
process) but instead solve this on your own application level. Using
Hashie is one example but another one, one I consider far better, is to
use your own classes. Consider the following:

 hash = {'id' => '123-abc-456', 'body' => 'hello'}

 if hash['id'] and !hash['id'].empty?
   puts "Processing message ID #{hash['id']}"
 end

 if hash['body'] and !hash['body'].empty?
   do_something(hash['body'])
 end

This is not a flaw in your proposal in particular but it's one of the
reasons why I'm not a big fan of using Hashes all over the place. If in
this example the "id" key is suddenly changed to "ID" you now have at
least 3 places where you have to modify your code and most likely other
places further down the line. This can be solved by doing something as
simple as the following:

 class Message
   attr_reader :id, :body

   def initialize(options)
     @id   = options['id']
     @body = options['body']
   end

   def has_id?
     return @id && !@id.empty?
   end

   def has_body?
     return @body && !@body.empty?
   end
 end

This allows you to write the following instead:

 message = Message.new({'id' => '123-abc-456', 'body' => 'hello'})

 if message.has_id?
   puts "Processing message ID #{message.id}"
 end

 if has_body?
   do_something(message.body)
 end

In this specific example it may seem a bit like an overkill but if that
Hash gets used in a dozen places you can save yourself tremendous
amounts of time by just wrapping a class around it.

Yorick

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 06-02-2013 16:22, Yorick Peterse escreveu:

What I'm trying to say is that the main reason why symbols exist in
Ruby in the first place is performance from what I've been told.
Correct, and your proposed changes would completely nullify those
performance benefits (see below).

People reading some Ruby book will notice that it is not particularly
designed with performance in mind but it is designed mostly towards
programmer's happiness. If that is the case, then worrying about
bothered programmers makes sense to a language like Ruby in my
opinion.
So basically what you're saying is "Ruby is written for happiness and
not performance, lets make it even more slow!". I'd rather see a world
where Ruby is both fast (enough) and easy to use instead of it being
easy to use and slower than a sloth.

Regarding the benchmarking information, you're missing a crucial aspect.
While the numbers in the specific examples I gave both clearly show that
the use of Strings is substantially slower. Yes, it's "only" 112 kb but
the difference will keep growing and growing until you hit your memory
limit.

Man, you're instantiating 50 millions strings and it only increased the
memory in 112KB. If your application creates so many strings that won't
be garbage collected then it is unlikely that symbols would help as a
replacement.

And "growing until you hit your memory limit" is actually only valid for
symbols, not for strings that are garbage collected already. Unless you
have some leak in your code that prevent those strings from being
collected by GC.

This is exactly one of the reasons Symbols exist: to make it easier and
faster to use commonly re-used Strings. The best example of this are
Hash keys.

Most of the programming languages don't support the concept of symbols
like Ruby. And you won't see C or C++ programmers complaining about this
neither.

This isn't possible when you're serializing/deserializing using some
library like JSON or any other. You don't control how hashes are
created by such libraries.
Of course it is. Marshal allows you to store arbitrary Ruby objects
(with the exception of a few such as Proc instances), in other cases you
can re-create your objects based on the supplied Hash.

Marshal is not portable across multiple languages (I use both Groovy and
Ruby in my overall application interacting with Redis). I'm talking
about JSON here. You don't have to find an alternative to JSON. Just try
to understand the issue I'm talking about.,

If you do not like using raw Hashes the solution in my opinion is not to
more or less re-write Ruby (and break everything that exists in the
process)

Like what?

but instead solve this on your own application level. Using
Hashie is one example but another one, one I consider far better, is to
use your own classes. Consider the following:

Ok, I won't repeat myself. Please give an example for the Redis + JSON
serialization use case presented in the ticket description.

Otherwise you cleared missed the point.

Updated by yorickpeterse (Yorick Peterse) about 11 years ago

And "growing until you hit your memory limit" is actually only valid
for symbols, not for strings that are garbage collected already.
Unless you have some leak in your code that prevent those strings from
being collected by GC.

Since existing code (and developers) assume that Symbols are only
created once they are generally used all over the place without any side
effects. The moment you start garbage collecting them there's an
increased chance of the GC kicking in right in the middle of (say) an
HTTP request. Yes, you may now be able to use both Symbols and Strings
as Hash keys but you now have to deal with increased GC activity.

Note that this of course depends on the code you're using. If you use
carefully written code that doesn't use Symbols this is not going to be
an issue. However, pretty every single Gem out there uses them and a lot
of them also use them quite heavily.

Most of the programming languages don't support the concept of symbols
like Ruby. And you won't see C or C++ programmers complaining about
this neither

C has a goto operator, does that mean Ruby should have one too (I'm
aware it's already there, it's not just enabled unless you specify some
compiler flag)? Just because one language has feature X it doesn't mean
all the others should have it too.

Marshal is not portable across multiple languages (I use both Groovy
and Ruby in my overall application interacting with Redis). I'm
talking about JSON here. You don't have to find an alternative to
JSON. Just try to understand the issue I'm talking about.,

It wasn't suggested as an alternative, merely an example that there is a
way of serializing arbitrary Ruby data.

Ok, I won't repeat myself. Please give an example for the Redis +
JSON serialization use case presented in the ticket description.

Otherwise you cleared missed the point.

Take a closer look at my previous Email, there's a fairly big example at
the bottom of it that you can't really miss. However, just in case:

 require 'redis'
 require 'json'

 client = Redis.new

 client.set('user', JSON({'id' => 1, 'name' => 'John Doe'}))

 # This would happen somewhere else (e.g. an external process)
 hash = JSON(client.get('user'))
 user = User.new(hash)

 # instead of user['id'] you can now just do `user.id` which means
 # that if the key name ever changes you only have to change it in
 # one place.
 if user.id
   # ...
 end

Another benefit is that this lets you attach your own methods to the
object without having to monkeypatch existing classes or using helper
methods (which feels very much like procedural programming).

Yorick

Updated by drbrain (Eric Hodel) about 11 years ago

  • Status changed from Open to Rejected

=begin
This proposal has no description of how to overlay the functionality of strings (mutable) with symbols (immutable).

This was previously tried during 1.9 which had such a plan but was ultimately rejected.

Due to a previous attempt and failure along with the lack of a concrete plan in this feature request I will reject this.

As to symbols or strings as hash keys, you should almost always use strings. This is the current community best-practice consensus.

You should only use symbols as keys in a Hash if you have a small, fixed set of keys and do not use user input to look up items in the hash.

Converting a string to a symbol you look up in a hash is not recommended. You have created two hash lookups out of one (the first for string to symbol mapping, the second for the lookup in your hash) and you risk a DoS as symbols are not garbage collected.

Consider this benchmark:

require 'benchmark'

N = 10_000_000

Benchmark.bmbm do |bm|
bm.report 'NULL' do
h = { 'foo' => 'bar'}

  N.times do
    # null
  end
end

bm.report 'string key, string lookup' do
  h = { 'foo' => 'bar' }

  N.times do
    h['foo']
  end
end

bm.report 'symbol key, symbol lookup' do
  h = { :foo => 'bar' }

  N.times do
    h[:foo]
  end
end

bm.report 'symbol key, string intern lookup' do
  h = { :foo => 'bar' }

  N.times do
    h['foo'.intern]
  end
end

end

Here are the results:

Rehearsal --------------------------------------------------------------------
NULL 0.440000 0.000000 0.440000 ( 0.448186)
string key, string lookup 1.870000 0.000000 1.870000 ( 1.866935)
symbol key, symbol lookup 0.660000 0.000000 0.660000 ( 0.661466)
symbol key, string intern lookup 2.230000 0.000000 2.230000 ( 2.222772)
----------------------------------------------------------- total: 5.200000sec

                                     user     system      total        real

NULL 0.430000 0.000000 0.430000 ( 0.434306)
string key, string lookup 1.860000 0.000000 1.860000 ( 1.862942)
symbol key, symbol lookup 0.680000 0.000000 0.680000 ( 0.681767)
symbol key, string intern lookup 2.270000 0.010000 2.280000 ( 2.264888)

While symbol key and symbol lookup is 2.7 times faster than string key and string lookup, it is also 1.2 times faster than interning a string for a symbol lookup.

=end

Updated by phluid61 (Matthew Kerwin) about 11 years ago

There's another issue here, which has been overlooked because it's philosophical rather than technical:

symbols aren't strings.

A string's value is a sequence of characters, which can be iterated over and operated on.

A symbol's only value is itself, is used as a token, and only supports comparison and casting operations. In MRI it's actually implemented as an integer, which can be looked up by its name (like a C enum).

I strongly believe there is already too much conflation between the two (completely unrelated) types. We can never benefit by blurring them any further.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

@drbrain (Eric Hodel), that means I'm unable to do things like

results << {id: id, name: name}

and have to use the old syntax all the time (and worrying about the differences all the time):

results << {'id' => id, 'name' => 'name'}

I still believe MRI should try to optimize frozen strings instead of symbols and just make symbols behave like frozen strings. :symbol would be a shortcut to 'symbol'.freeze.

I don't really think any of those micro-benchmarks would justify all the hassle and unhappiness that symbols bring to Ruby programmers.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 06-02-2013 17:12, Yorick Peterse escreveu:

And "growing until you hit your memory limit" is actually only valid
for symbols, not for strings that are garbage collected already.
Unless you have some leak in your code that prevent those strings from
being collected by GC.
Since existing code (and developers) assume that Symbols are only
created once they are generally used all over the place without any side
effects. The moment you start garbage collecting them there's an
increased chance of the GC kicking in right in the middle of (say) an
HTTP request. Yes, you may now be able to use both Symbols and Strings
as Hash keys but you now have to deal with increased GC activity.

You currently already don't control when GC activity begins, so I don't
understand how this could be considered any disadvantage...

Note that this of course depends on the code you're using. If you use
carefully written code that doesn't use Symbols this is not going to be
an issue. However, pretty every single Gem out there uses them and a lot
of them also use them quite heavily.

Most of the programming languages don't support the concept of symbols
like Ruby. And you won't see C or C++ programmers complaining about
this neither
C has a goto operator, does that mean Ruby should have one too (I'm
aware it's already there, it's not just enabled unless you specify some
compiler flag)? Just because one language has feature X it doesn't mean
all the others should have it too.

I didn't mean to say that Ruby should take inspiration on other
languages. My only intent was to show that symbols are not really
required. But then I remembered that both C, C++ and Java support
constant strings. In that sense Ruby could just optimize symbols to
behave like frozen strings and try to optimize that just like the other
languages optimize their constants.

Marshal is not portable across multiple languages (I use both Groovy
and Ruby in my overall application interacting with Redis). I'm
talking about JSON here. You don't have to find an alternative to
JSON. Just try to understand the issue I'm talking about.,
It wasn't suggested as an alternative, merely an example that there is a
way of serializing arbitrary Ruby data.

I know there are marshaling libraries in Ruby. I just don't understand
how that is relevant to this ticket.

Ok, I won't repeat myself. Please give an example for the Redis +
JSON serialization use case presented in the ticket description.

Otherwise you cleared missed the point.
Take a closer look at my previous Email, there's a fairly big example at
the bottom of it that you can't really miss.

I didn't miss it. I just didn't keep it in the reply because it is not
relevant to my use case.

However, just in case:

 require 'redis'
 require 'json'

 client = Redis.new

 client.set('user', JSON({'id' =>  1, 'name' =>  'John Doe'}))

 # This would happen somewhere else (e.g. an external process)
 hash = JSON(client.get('user'))
 user = User.new(hash)

 # instead of user['id'] you can now just do `user.id` which means
 # that if the key name ever changes you only have to change it in
 # one place.
 if user.id
   # ...
 end

Ok, you missed the point. Let me show you a complete example as I think
it will help you understanding my concerns (I thought I made myself
clear in the ticket description, but I hope this code example will help
you understand what I meant):

require 'redis'
require 'json'
require 'sequel'

cache = Redis.new
users = cache['users'] || begin
db = Sequel.connect('postgres://user:password@localhost/mydb')
cache['users'] = db[:users].select(:id, :name).map{|r| id: r[:id],
name: r[:name]} # or just select(:id, :name).all
end

p users.first[:id]

What will be the output of the code above?

Exactly! It depends! If the results were cached it will print "nil",
otherwise it will print "1" (or whatever the first id is).

This is the problem that symbols cause because they don't behave like
strings.

Best,
Rodrigo.

Updated by david_macmahon (David MacMahon) about 11 years ago

Hi, Rodrigo,

FWIW, I sympathize with your symbols-vs-strings as keys frustration, but I think it's not so trivial to have the best of both worlds (flexibility and performance). Here is a simplification of your example:

require 'redis'
cache = Redis.new
value = :value
key = 'key'

v1 = cache[key] || (cache[key] = value)
v2 = cache[key] || (cache[key] = value)

p v1 # :value or "value"
p v2 # always "value"
p cache[key] == value # always false

IMHO, the crux of the problem here is that Redis converts all Symbols to Strings, but assigning "v1 = cache[key] = value" does not store a String in v1 if value is a Symbol. This could be addressed by doing:

v1 = cache[key] || (cache[key] = value; cache[key])

which will store a String in v1 if value is a Symbol even if key does not yet exist in cache. Yes, it is an extra cache fetch.

The same kind of thing happens if value is a Fixnum. Surely you wouldn't want Fixnums and Strings to be the same thing! :-)

Thanks for starting this fascinating thread,
Dave

On Feb 6, 2013, at 3:45 PM, Rodrigo Rosenfeld Rosas wrote:

Ok, you missed the point. Let me show you a complete example as I think it will help you understanding my concerns (I thought I made myself clear in the ticket description, but I hope this code example will help you understand what I meant):

require 'redis'
require 'json'
require 'sequel'

cache = Redis.new
users = cache['users'] || begin
db = Sequel.connect('postgres://user:password@localhost/mydb')
cache['users'] = db[:users].select(:id, :name).map{|r| id: r[:id], name: r[:name]} # or just select(:id, :name).all
end

p users.first[:id]

What will be the output of the code above?

Exactly! It depends! If the results were cached it will print "nil", otherwise it will print "1" (or whatever the first id is).

This is the problem that symbols cause because they don't behave like strings.

Best,
Rodrigo.

Updated by phluid61 (Matthew Kerwin) about 11 years ago

rosenfeld (Rodrigo Rosenfeld Rosas) wrote:

@drbrain (Eric Hodel), that means I'm unable to do things like

results << {id: id, name: name}

and have to use the old syntax all the time (and worrying about the differences all the time):

results << {'id' => id, 'name' => 'name'}

You say "old syntax," maybe you should think of it as the "general syntax." Or, as I think of it, the "real syntax." The new {id: id} syntax is special case sugar for Hashes keyed on Symbols, which we've already determined is not a great thing to be doing in most cases.

I still believe MRI should try to optimize frozen strings instead of symbols and just make symbols behave like frozen strings. :symbol would be a shortcut to 'symbol'.freeze.

But symbols aren't strings. You can't do any of the string-specific things on them, except by first casting them to strings. Why create a string, then use it to generate an (unrelated) type of object? Just create the symbol in the first place (as already happens).

I don't really think any of those micro-benchmarks would justify all the hassle and unhappiness that symbols bring to Ruby programmers.

A simple solution presents itself: stop using them. Or I should say, stop using them except when you absolutely must.

If we stop thinking of them as special/frozen/whatever strings, and stop using them as "strings with construction optimisation," this whole issue becomes irrelevant. Then we can also stop using the symbol-specific {a: a} syntax, except when it really makes sense. And voila! everything is gravy.

Earlier you also wrote:

You can't know in advance if a hash returned by some external code is indexed by string or symbols. You have to test by yourself or check the documentation.

Indeed. Similarly you can't know in advanced if it's keyed on integers, or CustomMagicClass instances. Not a symbol-specific issue. If the library authors cleared up their misconception that symbols are magic strings, you wouldn't have a problem. Solution: work around their shortcomings, or find/write a better library.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

phluid61 (Matthew Kerwin) wrote:

You say "old syntax," maybe you should think of it as the "general syntax." Or, as I think of it, the "real syntax." The new {id: id} syntax is special case sugar for Hashes keyed on Symbols, which we've already determined is not a great thing to be doing in most cases.

I agree that a string is what I want in all cases. That is exactly why I don't feel the need for symbols. If symbols are just really required as a fundamental implementation detail of the MRI implementation, then I don't think it is a good reason to justify keeping them in the language level. Just find other ways to optimize methods/etc lookup in the internal MRI code. This should be a separate discussion from the language design itself.

I'd really prefer you to think if symbols are really a good thing to have in the design of the Ruby language if you forget about all performance impacts it might have on the MRI implementation details. Then, still forgetting about performance and internal implementation details, try to reason why :symbol != 'symbol' is useful in Ruby just like a[:a] != a['a']. I've been using Ruby for several years now and I can tell you for sure that people often want them to behave the same and they don't want to worry about performance impact either. People just don't know when to use symbols and strings.

Take the Sequel library for instance.

DB[:user].select(:id, :name).all will return [{id: 1, name: 'Rodrigo'}] as opposed to [{'id' => 1}, {'name' => 'Rodrigo'}]

A similar case happens all around and people simply are confused whether they should be using symbols or strings in their code. This is specially misleading in hashes, specially because it is not uncommon that you can intermix strings and symbols as hash keys causing frequent bugs. It doesn't make sense to argue vs String vs Integer or Symbol vs Integer because this kind of bug doesn't really happen. People don't confuse Integer with Strings or Symbols. They confuse Symbols with Strings only. And each library will use a different criteria to decide if symbols or strings should be used and that forces each programmer to worry about those differences between libraries.

I still believe MRI should try to optimize frozen strings instead of symbols and just make symbols behave like frozen strings. :symbol would be a shortcut to 'symbol'.freeze.

But symbols aren't strings.

I know they aren't. That is why I'm asking to change this behavior!

You can't do any of the string-specific things on them, except by first casting them to strings.

We all know how symbols are different from strings, it doesn't help repeating it all the way. I'd prefer that you focus on explaining why you think keeping symbols a separate beast is of any usefulness (ignore any performance concerns for now, I'd like to make sure performance is the only factor involved here first).

Why create a string, then use it to generate an (unrelated) type of object? Just create the symbol in the first place (as already happens).

This is what I do but I don't control other libraries. Anyway this makes the new sexy hash syntax almost unuseful to me since strings is what I want most of the times. And I really do hate the general syntax for hashes. The new one is more compact and takes me much less type to type and it is also similar to what most languages do (JavaScript, Groovy, etc). The difference is that in the other languages a string is used since they don't have the symbols concept.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Also, so that you stop arguing that the differences between symbols and strings are just like the differences between strings and integers (non-sense), notice that HashWithIndifferentAccess makes this distinction:

h = HashWithIndifferentAccess.new({1 => 'a', a: 'a'})
h['1'] == nil
h[1] == 'a'
h['a'] == 'a'
h[:a] == 'a'

Since you don't see any popular hash implementation that will consider h[1] == h['1'] (like JavaScript), you could take the conclusion that people only really care that string behave differently than symbols.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

rosenfeld (Rodrigo Rosenfeld Rosas) wrote:

cache = Redis.new
users = cache['users'] || begin
db = Sequel.connect('postgres://user:password@localhost/mydb')
cache['users'] = db[:users].select(:id, :name).map{|r| id: r[:id],
name: r[:name]} # or just select(:id, :name).all

Sorry, I forgot the JSON conversion in the example above:

cache = Redis.new
db = Sequel.connect('postgres://user:password@localhost/mydb')
users = if cached = cache['users']
JSON.parse cached
else
db[:users].select(:id, :name).all.tap{|u| cache['users'] = JSON.unparse u}
end

I know the code above could be simplified, but I'm avoiding the parse -> unparse operation when it is not required (although it would make the code always work in this case):

users = JSON.parse (cache['users'] ||= JSON.unparse db[:users].select(:id, :name).all)

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

david_macmahon (David MacMahon) wrote:

Hi, Rodrigo,

FWIW, I sympathize with your symbols-vs-strings as keys frustration, but I think it's not so trivial to have the best of both worlds (flexibility and performance).

I'm not asking for performance nor flexibility. The only reason I didn't suggest to simply remove symbols at all is because I know it would be promptly rejected since all Ruby code would have to be changed to accomplish a change like that. And I don't really believe that symbols help the overall performance of any Ruby program. I feel it is more likely to reduce performance because people are often using conversions between symbols and strings making the overall performance slower, not faster.

...
IMHO, the crux of the problem here is that Redis converts all Symbols to Strings...

Sorry, my example was wrong, I forgot about the JSON serialization. I don't really know what the Redis library does if I try to store a non-string object so I never tried to do so. I'm trying to avoid an unnecessary JSON.parse operation when I don't need it. See my comments above.

Updated by phluid61 (Matthew Kerwin) about 11 years ago

On 7 February 2013 20:46, rosenfeld (Rodrigo Rosenfeld Rosas) wrote:

I agree that a string is what I want in all cases. That is exactly why I
don't feel the need for symbols. If symbols are just really required as
a fundamental implementation detail of the MRI implementation, then I
don't think it is a good reason to justify keeping them in the language
level. Just find other ways to optimize methods/etc lookup in the
internal MRI code. This should be a separate discussion from the language
design itself.

I'd really prefer you to think if symbols are really a good thing to
have in the design of the Ruby language if you forget about all
performance impacts it might have on the MRI implementation details.

Ok, methods. They have a bucket of queriable information (a Method
instance), and they have a symbolic representation (a Symbol). I don't
want to have to instantiate an entire Method object (or a whole bunch of
them) every time I want to talk to an object abouts its methods; I just
want a single, simple, universal token that represents that (or those)
method(s).

Sorry, that's a performance optimisation detail. Ok, I don't want to have
to instantiate a Method object that potentially doesn't have a
corresponding method. That could be confusing.

You will now argue that I could as easily use a String as a Symbol, and
yes, ignoring performance and implementation details that is true. But I
don't want to write code that performs poorly. If, in thise case, exposing
implementation details make my code easier and better, then by the gods,
let me use it. It is then up to me not to misuse it. Similarly: why have
any numeric class that isn't Rational?

And for the record: "I don't ever want to use ClassX so let's remove it"
is, frankly, silly.

Then, still forgetting about performance and internal implementation
details, try to reason why :symbol != 'symbol' is useful in Ruby just
like a[:a] != a['a']. I've been using Ruby for several years now and I
can tell you for sure that people often want them to behave the same and
they don't want to worry about performance impact either.

Ok, completely philosphically, without any reference to performance or
implementation details, why is a Java enum not equivalent to (or auto-cast
to and from) a Java string? An enum is just a token, yes? It looks like a
string; it often spells a word that people might like to read, even
capitalise. But it's not a string. It's something else. A Symbol is
exactly like that enum.

I, too, have been using Ruby for several years now; and I, too, have seen a
lot of people wanting Symbol and String to behave the same. Hells, at
times even I have wanted that. But the simple fact is: those people
(myself included) are wrong. If they want a String, use a String. If they
want to force a Symbol-shaped peg into a String-shaped hole, then they'll
have to do whatever hoop-jumping is required; exactly as if you want a Java
enum to support implicit coercion to and from a string.

People just don't know when to use symbols and strings.

Bingo. Your solution is: hide Symbols from those people. My solution is:
don't change anything; maybe eventually enough people will learn that the
two classes are, in fact, different.

Take the Sequel library for instance.

No thanks, apparently the authors don't know the difference between Symbols
and Strings.

We all know how symbols are different from strings,

Well apparently not, otherwise this would be a non-issue.

it doesn't help repeating it all the way.

Perhaps I believe that if I say it enough times, in enough places, people
might actually notice. And maybe even listen.

I'd prefer that you focus on explaining why you think keeping symbols a
separate beast is of any usefulness

I'll choose to interpret that as "... why I think keeping symbols at all
...". Simply: because they're already here. Relegating them to an
implementation detail and hiding them from the language will only break
100% of existing code. Some of that code is good code. Is it worth
breaking everything so a bunch of people can't accidentally use ClassX when
they should be using ClassY?

-- I'll inject this later comment here, because it's topical:

Also, so that you stop arguing that the differences between symbols and
strings are just like the differences between strings and integers
(non-sense), notice that HashWithIndifferentAccess makes this distinction:
[...]
Since you don't see any popular hash implementation that will consider
h[1] == h['1'] (like JavaScript), you could take the conclusion that
people only really care that string behave differently than symbols.

Yes, but those are people who don't know the difference between Symbols and
Strings. Just because they don't know it, doesn't make it untrue.
Personally I've never used HashWithIndifferentAccess, or needed to.

Incidentally those people don't want a Hash at all. They want an
associative array, one that uses something like <=> or === to compare keys
(instead of #hash and #eql?). If RBTree was more mature,
HashWithIndifferentAccess wouldn't be needed. Shall we repeat this
discussion, but this time about Hash and assoc.array instead of Symbol and
String?

This is what I do but I don't control other libraries.

This is true of any issue in a library. If you think the library's
benefits outweigh its costs, then you use the library. If the fact that
the authors erroneously conflate Symbols and Strings is outweighed by the
fact that it's otherwise extremely useful, it's up to you to work around
the shortcomings. Just like if some otherwise brilliant library uses 0
instead of nil, or something.

Anyway this makes the new sexy hash syntax almost unuseful to me since
strings is what I want most of the times.

So, like I said before, just don't use it.

And I really do hate the general syntax for hashes. The new one is more
compact and takes me much less type to type and it is also similar to
what most languages do (JavaScript, Groovy, etc).

The general syntax served us well enough through 1.8 and 1.9. Personally I
preferred being able to use : at the end of if/when/etc. statements. If
I want to use javascript syntax, there's always node.js

The difference is that in the other languages a string is used since
they don't have the symbols concept.

That's a good point. I'd love to be able to change the new syntax so {a:1}
meant {'a'=>1}, but that's not going to happen. As such, in your eyes and
mine, the new syntax is useless for most intents and purposes, so we might
as well keep on writing Ruby code the way we always have (with => tokens).

P.S. sorry to anyone else who is sick of this conversation, but I think it
needs to be had. Let us know if we should take it offline somewhere.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 07-02-2013 10:04, Matthew Kerwin escreveu:

On 7 February 2013 20:46, rosenfeld (Rodrigo Rosenfeld Rosas) wrote:

I agree that a string is what I want in all cases. That is exactly why I
don't feel the need for symbols. If symbols are just really required as
a fundamental implementation detail of the MRI implementation, then I
don't think it is a good reason to justify keeping them in the language
level. Just find other ways to optimize methods/etc lookup in the
internal MRI code. This should be a separate discussion from the
language
design itself.

I'd really prefer you to think if symbols are really a good thing to
have in the design of the Ruby language if you forget about all
performance impacts it might have on the MRI implementation details.

Ok, methods. They have a bucket of queriable information (a Method
instance), and they have a symbolic representation (a Symbol). I
don't want to have to instantiate an entire Method object (or a whole
bunch of them) every time I want to talk to an object abouts its
methods; I just want a single, simple, universal token that represents
that (or those) method(s).

Like a string?

Sorry, that's a performance optimisation detail.

Before I continue with my arguments and could focus solely on the
performance issue I'd like to confirm that there is no other reason why
symbols do exist. If performance is the sole reason and if I can make
the point that symbols actually degrades most Ruby programs rather than
improve the overall performance then I can still maintain my hopes for
this ticket.

...
And for the record: "I don't ever want to use ClassX so let's remove
it" is, frankly, silly.

This is clearly not the reason here. That's why I'm asking: what Symbols
are useful for? Is it performance the only reason why symbols exist?

Symbols cause lots of confusion as I showed in previous examples. That's
why I want to remove it, but I didn't ask to remove it in this ticket
anyway. Just to make it behave exactly like strings.

...Ok, completely philosphically, without any reference to performance
or implementation details, why is a Java enum not equivalent to (or
auto-cast to and from) a Java string?

Java, C and C++ have different goals then Ruby. They aim at the best
possible performance given their constraints. They are also statically
typed. Enums have two goals in such languages. Improving performance and
reducing memory footprint is one of them. The other one is to help the
compiler to find errors at compile time by restricting the input type in
some functions/methods and variables. I don't really understand how this
is relevant to this discussion.

I, too, have been using Ruby for several years now; and I, too, have
seen a lot of people wanting Symbol and String to behave the same.
Hells, at times even I have wanted that. But the simple fact is:
those people (myself included) are wrong. If they want a String, use
a String. If they want to force a Symbol-shaped peg into a
String-shaped hole, then they'll have to do whatever hoop-jumping is
required; exactly as if you want a Java enum to support implicit
coercion to and from a string.

I don't want that for Java enums and I don't really understand how Java
enums relate to the string vs symbols debate in Ruby.

People just don't know when to use symbols and strings.

Bingo. Your solution is: hide Symbols from those people.

Yes!

My solution is: don't change anything; maybe eventually enough people
will learn that the two classes are, in fact, different.

They won't.

Take the Sequel library for instance.

No thanks, apparently the authors don't know the difference between
Symbols and Strings.

But I really love the library. Should I really stop using it just
because it returns an array of hashes indexed by symbols? And Sequel is
not the only library doing so. Should I stop using all gems out there
because the authors don't understand they should be always using strings
instead of symbols in such cases?

You should ask yourself: why are authors so confusing about whether to
use strings or symbols? Are they all just stupid? Isn't it clear in all
Ruby books? No, it isn't! It is just really confusing. I'm yet to read
some book that does a strong argument whether you should be using
symbols or strings. They just say that symbols perform better than
strings so authors think: "hey, then I'll just use symbols everywhere
and my gems will perform the best possible way!". But this thinking is
plain wrong because you'll need extra steps for conversions among those
types very often. The fact is that most authors don't really care about
symbols or strings at all. They don't spend their time thinking about
whether they should be using symbols or strings. They don't WANT to
worry about it! And they're not wrong! Since they don't want someone
looking at their code and telling them that their gem could perform
better if they used symbols instead of strings they will just use
symbols everywhere! This reality won't change. That is why I think
programmers shouldn't have to worry about any performance difference
that might exist between using symbols or strings in their code.

If there is a real boost using symbols internally in MRI then this
should be an implementation detail only, not exposed to Ruby programs.
That is why I suggested the optimizations to care if the string is
frozen or not (like other compilers will optimize constants) instead of
creating a new concept (symbols) just for that. They could keep the
:symbol syntax as an alias to 'symbol'.freeze.

...

I'd prefer that you focus on explaining why you think keeping symbols a
separate beast is of any usefulness

I'll choose to interpret that as "... why I think keeping symbols at
all ...". Simply: because they're already here.

This is not a good argument in my opinion. If you want to keep the
syntax :name as an alias to 'name'.freeze I believe most current Ruby
programs wouldn't be affected by such change.

... Personally I've never used HashWithIndifferentAccess, or needed to.

Me neither. But for different reasons. I need the behavior but I don't
think it worths the extra dependencies in my code (ActiveSupport) nor
the cumbersome of writing such a big class name everytime I want my hash
to behave like HWIA. I prefer to take some time to investigate if all my
hash keys are really strings than to just instantiate HWIA all over the
places.

Incidentally those people don't want a Hash at all. They want an
associative array, one that uses something like <=> or === to compare
keys (instead of #hash and #eql?).

:a === 'a' is not true so I don't understand how such suggested
associative array would help here.

...
This is true of any issue in a library. If you think the library's
benefits outweigh its costs, then you use the library. If the fact
that the authors erroneously conflate Symbols and Strings is
outweighed by the fact that it's otherwise extremely useful, it's up
to you to work around the shortcomings. Just like if some otherwise
brilliant library uses 0 instead of nil, or something.

Again, please don't pretend that the confusion between strings and
symbols are similar to confusions between 0 and nil.

The general syntax served us well enough through 1.8 and 1.9.

Actually I never liked writing hashes as {key => value} in all years
I've been working with Ruby. But I won't stop using Ruby just because I
don't like its hash declaration syntax just the way I won't replace
Sequel just because they return hashes indexed by symbols instead of
strings.

...

The difference is that in the other languages a string is used since
they don't have the symbols concept.

That's a good point. I'd love to be able to change the new syntax so
{a:1} meant {'a'=>1}, but that's not going to happen.

I agree it is unlike to happen. What about another syntax: {{a: 1}} =>
{'a' => 1}? Maybe it would worth trying to ask for some syntax change
like this one. We could even add interpolation to it: {{"value
#{computed}": 1}}.

Updated by jeremyevans0 (Jeremy Evans) about 11 years ago

phluid61 (Matthew Kerwin) wrote:

Take the Sequel library for instance.

No thanks, apparently the authors don't know the difference between Symbols
and Strings.

Sequel uses symbol keys instead of string keys intentionally. Sequel maps SQL features directly to ruby objects, mapping SQL identifiers (columns/tables/aliases) to ruby symbols and SQL strings to ruby strings. SQL query results can be thought of as a mapping of SQL identifiers to the values for each identifier in the query, thus Sequel uses a hash with symbol keys.

Ruby uses symbols in a very similar way to how SQL uses identifiers, with symbols basically acting as an identifier. The fact that ruby uses symbols as identifiers should be obvious to anyone who has looked at MRI's implementation, where a symbol is simply an alternate representation of an ID (the internal identifier type that MRI uses to map names to values):

#define ID2SYM(x) (((VALUE)(x)<<RUBY_SPECIAL_SHIFT)|SYMBOL_FLAG)
#define SYM2ID(x) RSHIFT((unsigned long)(x),RUBY_SPECIAL_SHIFT)

The basic philosophical difference between a symbol and a string is a ruby symbol is an identifier, while a ruby string just represents arbitrary data. You should convert a string to a symbol if you know that the data the string contains represents an identifier you would like to use. You should convert a symbol to a string if you are using name of the identifier as data.

Updated by david_macmahon (David MacMahon) about 11 years ago

On Feb 7, 2013, at 3:00 AM, rosenfeld (Rodrigo Rosenfeld Rosas) wrote:

Issue #7792 has been updated by rosenfeld (Rodrigo Rosenfeld Rosas).

rosenfeld (Rodrigo Rosenfeld Rosas) wrote:

cache = Redis.new
users = cache['users'] || begin
db = Sequel.connect('postgres://user:password@localhost/mydb')
cache['users'] = db[:users].select(:id, :name).map{|r| id: r[:id],
name: r[:name]} # or just select(:id, :name).all

Sorry, I forgot the JSON conversion in the example above:

cache = Redis.new
db = Sequel.connect('postgres://user:password@localhost/mydb')
users = if cached = cache['users']
JSON.parse cached
else
db[:users].select(:id, :name).all.tap{|u| cache['users'] = JSON.unparse u}
end

I still think the fundamental issue is that Sequel is returning something that contains symbols, but JSON.parse(JSON.unparse(x)) will never return anything containing a symbol even if x does. Because of this the "users" variable is assigned something that contains symbols in one case and something that contains no symbols in all other cases. IMHO, the example shows not a deficiency in the Ruby language itself but rather a (subtle, easy to fall victim to) coding pitfall of not properly managing/reconciling the incompatible return types from the two libraries.

I know the code above could be simplified, but I'm avoid the parse -> unparse operation when it is not required (although it would make the code always work in this case):

users = JSON.parse (cache['users'] ||= JSON.unparse db[:users].select(:id, :name).all)

This seems an excellent solution IMHO as it clearly results in "users" being assigned the return value of JSON.parse every time and only results in one extra JSON.parse call only on the first time through. That seems a pretty small price to pay for reconciling the two libraries' incompatible return return types.

Dave

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

david_macmahon (David MacMahon) wrote:

I still think the fundamental issue is that Sequel is returning something that contains symbols, but JSON.parse(JSON.unparse(x)) will never return anything containing a symbol even if x does. Because of this the "users" variable is assigned something that contains symbols in one case and something that contains no symbols in all other cases. IMHO, the example shows not a deficiency in the Ruby language itself but rather a (subtle, easy to fall victim to) coding pitfall of not properly managing/reconciling the incompatible return types from the two libraries.

But don't you agree that we wouldn't have this kind of problem if Symbols behave just like Strings? Or if Symbols didn't exist at all? Or even if hashes always behaved as HWIA? This is what I'm talking about. But the biggest problem is that I can't really see any real advantage on symbols behaving differently than strings, except maybe for internal usage of symbols.

I know the code above could be simplified, but I'm avoid the parse -> unparse operation when it is not required (although it would make the code always work in this case):

users = JSON.parse (cache['users'] ||= JSON.unparse db[:users].select(:id, :name).all)

This seems an excellent solution IMHO as it clearly results in "users" being assigned the return value of JSON.parse every time and only results in one extra JSON.parse call only on the first time through. That seems a pretty small price to pay for reconciling the two libraries' incompatible return return types.

You seem to be underestimating how slow JSON.parse can be on big lists. I've spent a lot of time in the past doing benchmarks for several JSON parsers implementations available for Ruby because most of the time spent on some requests were caused by JSON#unparse. JSON#unparse was the bottleneck of some requests. I don't think JSON#parse would be much faster. An extra parse operation on a big list could easily add 100ms to the request timing. That is why I want to avoid it.

Also, it doesn't happen only the first time but on each request.

Updated by david_macmahon (David MacMahon) about 11 years ago

On Feb 7, 2013, at 10:23 AM, rosenfeld (Rodrigo Rosenfeld Rosas) wrote:

Issue #7792 has been updated by rosenfeld (Rodrigo Rosenfeld Rosas).

david_macmahon (David MacMahon) wrote:

I still think the fundamental issue is that Sequel is returning something that contains symbols, but JSON.parse(JSON.unparse(x)) will never return anything containing a symbol even if x does. Because of this the "users" variable is assigned something that contains symbols in one case and something that contains no symbols in all other cases. IMHO, the example shows not a deficiency in the Ruby language itself but rather a (subtle, easy to fall victim to) coding pitfall of not properly managing/reconciling the incompatible return types from the two libraries.

But don't you agree that we wouldn't have this kind of problem if Symbols behave just like Strings? Or if Symbols didn't exist at all? Or even if hashes always behaved as HWIA?

Yes, I agree that we wouldn't have this kind of problem if any of those alternatives existed. I'm just not (yet) convinced that any of those alternatives are desirable just to avoid this kind of problem. I do somewhat like the idea of having Hash behave as HWIA. I think the number of real world uses of something like:

{:a => 0, 'a' => 1}

is likely to be very small.

I know the code above could be simplified, but I'm avoid the parse -> unparse operation when it is not required (although it would make the code always work in this case):

users = JSON.parse (cache['users'] ||= JSON.unparse db[:users].select(:id, :name).all)

This seems an excellent solution IMHO as it clearly results in "users" being assigned the return value of JSON.parse every time and only results in one extra JSON.parse call only on the first time through. That seems a pretty small price to pay for reconciling the two libraries' incompatible return return types.

An extra parse operation on a big list could easily add 100ms to the request timing.

So use explicit initialization instead of lazy initialization.

Also, it doesn't happen only the first time but on each request.

In the original example, JSON.parse was used on every call except the first one. The modified example uses JSON.parse on every call including the first one. That's why I said the modified example has only one extra JSON.parse call (i.e. the one extra one on the first call). If that's too much overhead, either don't use lazy initialization or explicitly invoke the method once at startup to force lazy initialization so it doesn't impact the (un?)lucky first user.

Dave

Updated by trans (Thomas Sawyer) about 11 years ago

I think the best thing to do about this is simply to ask that a HWIA class be added to Ruby's core. That's as far as this issue can possibly go at this time. And be very happy if it happens b/c 1) its hard to get such things to happen and 2) it's a nice big leap in the right direction. Asking for Hash to become a HWIA, is simply asking for too much backward compatibility breakage for a non-sufficient reason. Keeping on about it will do no good and simply fall on deaf ears, and rather then encourage the inclusion of a HWIA, may well do the opposite.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 07-02-2013 16:43, David MacMahon escreveu:

...An extra parse operation on a big list could easily add 100ms to the request timing.
So use explicit initialization instead of lazy initialization.

Sorry, didn't get. Could you please show some sample code?

Also, it doesn't happen only the first time but on each request.
In the original example, JSON.parse was used on every call except the first one. The modified example uses JSON.parse on every call including the first one. That's why I said the modified example has only one extra JSON.parse call (i.e. the one extra one on the first call). If that's too much overhead, either don't use lazy initialization or explicitly invoke the method once at startup to force lazy initialization so it doesn't impact the (un?)lucky first user.

I see the confusion. I simplified the code in that example. Here is how
it would look in a real Rails controller:

class MyController

def my_action
   @users = if cached = CacheStore.fetch('users')
     JSON.parse cached
   else
     DB[:users].select(:id, :name).all.tap{|u| CacheStore.store 

'users', JSON.unparse u}
end
end
end

Of course I don't cache the users list, this is just a small example.
The real query is much more complex and could take up to a second when
lots (20) of fields are searched using the user interface. Usually the
query would take about 100ms or less (up to 5 fields usually) but then
the user may want to print the results or export to Excel or changing to
another page and caching would help a lot in that case.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Thomas, if you think we could get some constructor for HWIA, like say, {{a: 'a'}}, then it would already help a lot. But I don't see typing such a long name (HashWithIndifferentAccess) whenever I need such behavior being it included on core or not.

Updated by david_macmahon (David MacMahon) about 11 years ago

On Feb 7, 2013, at 11:04 AM, Rodrigo Rosenfeld Rosas wrote:

Em 07-02-2013 16:43, David MacMahon escreveu:

...An extra parse operation on a big list could easily add 100ms to the request timing.
So use explicit initialization instead of lazy initialization.

Sorry, didn't get. Could you please show some sample code?

I was referring to the example, but now I see that the real-world variation is not simply lazy initialization, but rather caching of a result derived from user input that may or may not be used again. In the latter case, the result cannot be pre-initialized and an extra JSON.parse would be significant.

Sorry for veering off topic,
Dave

Updated by trans (Thomas Sawyer) about 11 years ago

@rosenfeld (Rodrigo Rosenfeld Rosas) Well I would hope for a much shorter name myself. e.g. Map, Index, Dict are some viable candidates.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

I could see myself using "Map(a: 1, b: 2)" instead of "{a: 1, b: 2}" when I want HWIA. Sounds good to you? Should I create a ticket like this one?

Updated by phluid61 (Matthew Kerwin) about 11 years ago

On 7 February 2013 23:09, Rodrigo Rosenfeld Rosas wrote:

Enums have two goals in such languages. Improving performance and
reducing memory footprint is one of them. The other one is to help the
compiler to find errors at compile time by restricting the input type in
some functions/methods and variables. I don't really understand how this
is relevant to this discussion.

No, no no no. Enums exist because they are identifiers. They are symbolic
representations of a concept that either does not necessarily otherwise
exist in the code, or does not have to be fully instantiated in order to
discuss it. That is exactly what Symbols are.

They don't spend their time thinking about whether they should be using
symbols or strings. They don't WANT to worry about it!

Your overarching goal is to coddle developers who write code without
understanding either the language, or the concepts behind their own code.
There is a massive philosophical disjunct here between you and I, and I
think it will never be overcome.

I agree it is unlike to happen. What about another syntax: {{a: 1}} =>
{'a' => 1}? Maybe it would worth trying to ask for some syntax change
like this one. We could even add interpolation to it:
{{"value #{computed}": 1}}.

You'd probably be more likely to succeed with a new %string-style notation,
like %h{a:1, b:2}. Although then again, possibly not.

Updated by phluid61 (Matthew Kerwin) about 11 years ago

On 8 February 2013 03:01, jeremyevans0 (Jeremy Evans) <
> wrote:

phluid61 (Matthew Kerwin) wrote:

Take the Sequel library for instance.

No thanks, apparently the authors don't know the difference between
Symbols
and Strings.

Sequel uses symbol keys instead of string keys intentionally. Sequel maps
SQL features directly to ruby objects, mapping SQL identifiers
(columns/tables/aliases) to ruby symbols and SQL strings to ruby strings.
SQL query results can be thought of as a mapping of SQL identifiers to the
values for each identifier in the query, thus Sequel uses a hash with
symbol keys.

Sorry, I was being glib.

Interestingly, this is actually an example of Symbols being used correctly
(or at least, not just out-right incorrectly). It's a pity people see this
is a problem to be worked around, rather than a feature.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 07-02-2013 19:11, Matthew Kerwin escreveu:

On 7 February 2013 23:09, Rodrigo Rosenfeld Rosas wrote:

Enums have two goals in such languages. Improving performance and
reducing memory footprint is one of them. The other one is to help the
compiler to find errors at compile time by restricting the input type in
some functions/methods and variables. I don't really understand how this
is relevant to this discussion.

No, no no no. Enums exist because they are identifiers. They are
symbolic representations of a concept that either does not necessarily
otherwise exist in the code, or does not have to be fully instantiated
in order to discuss it. That is exactly what Symbols are.

If you really believe symbols are similar to enums I guess you haven't
done much C, C++ or Java programming and used enums. Here is the main
reason why enums exist. First let me notice that C and Java implement
enums in different ways:

C example:

typedef enum {HEAD, TAIL} coin_side;
coin_side my_coin = HEAD; // my_coin = 0 would also work here

if you try to create another typedef like this, the compiler will
complain that HEAD is already declared:
typedef enum {HEAD, TAIL} alternate_coin_side;

if you do something like:

typedef enum {PAPER, ROCK, SCISORS} game;
coin_side my_coin = PAPER;

The C compiler won't complain. But Java takes a different approach when
it comes to enums:

class MyClass {
enum Game {PAPER, ROCK, SCISORS};
enum CoinSide {HEAD, TAIL};
void test(){
Game a = Game.PAPER;
Game b = CoinSide.HEAD; // won't compile!
}
}

In that sense, if you write a method accepting a coin side you can only
pass in a CoinSize enum. Not a number. Not another enum type. This is
what I understand by enums. Not symbols related at all in my opinion.

They don't spend their time thinking about whether they should be using
symbols or strings. They don't WANT to worry about it!

Your overarching goal is to coddle developers who write code without
understanding either the language, or the concepts behind their own
code. There is a massive philosophical disjunct here between you and
I, and I think it will never be overcome.

It is a matter of choosing the right tool. If you're really concerned
about some really small performance improvements you might get by using
symbols instead of strings I would question if Ruby is really the right
language for you.

I'd never consider Ruby or Java to write hard-real-time applications the
same way I wouldn't consider Windows or plain Linux for such task. I'd
most likely use C and Linux + Xenomai patch (if building for desktops)
instead.

When I'm writing Ruby code I don't want to worry about micro performance
improvements. The minimal amount of time I would probably care in
optimizing would be 100ms instead of micro-seconds when I'm writing C
programs that must complete some complex tasks in very short times when
writing real-time tasks.

I agree it is unlike to happen. What about another syntax: {{a: 1}} =>
{'a' => 1}? Maybe it would worth trying to ask for some syntax change
like this one. We could even add interpolation to it:
{{"value #{computed}": 1}}.

You'd probably be more likely to succeed with a new %string-style
notation, like %h{a:1, b:2}. Although then again, possibly not.

That's an idea, yes, but I guess I prefer Thomas' suggestion of using
Map(a: 1, b: 2).

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 07-02-2013 19:15, Matthew Kerwin escreveu:

On 8 February 2013 03:01, jeremyevans0 (Jeremy Evans)
< >
wrote:

phluid61 (Matthew Kerwin) wrote:
> > Take the Sequel library for instance.
>
>  No thanks, apparently the authors don't know the difference
between Symbols
>  and Strings.

Sequel uses symbol keys instead of string keys intentionally.
 Sequel maps SQL features directly to ruby objects, mapping SQL
identifiers (columns/tables/aliases) to ruby symbols and SQL
strings to ruby strings. SQL query results can be thought of as a
mapping of SQL identifiers to the values for each identifier in
the query, thus Sequel uses a hash with symbol keys.

Sorry, I was being glib.

Interestingly, this is actually an example of Symbols being used
correctly (or at least, not just out-right incorrectly). It's a pity
people see this is a problem to be worked around, rather than a feature.

I've never complained about Sequel returning symbols. I'm complaining
about symbols not behaving just like regular strings when they're used
as keys in hashes among other usage examples. That's why I never asked
Jeremy to change Sequel's current behavior.

Updated by phluid61 (Matthew Kerwin) about 11 years ago

Sent from my phone, so excuse the typos.
On Feb 8, 2013 8:50 AM, "Rodrigo Rosenfeld Rosas"
wrote:

If you really believe symbols are similar to enums I guess you haven't
done much C, C++ or Java programming and used enums. Here is the main
reason why enums exist.

Yes, the reason they are called 'enums' and not 'arbitrary identifiers' is
because they define an explicit, finite, enumerated set of identifiers; as
such an 'enum' is a set of things-that-are-like-symbols. But ´head´ is not,
and should never be, equal to 0 or 1 or whatever. The fact that C reveals
its implementation in this way is an artifact of C, not the concept of
enums in general.

Updated by Anonymous about 11 years ago

Hi,

During early stage of 1.9 development, I tried to make symbols and
strings behave same (at least similar), and it had broken too many
programs. I understand your problem but it's not worth raising huge
compatibility issues.

						matz.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 07-02-2013 21:58, Yukihiro Matsumoto escreveu:

Hi,

During early stage of 1.9 development, I tried to make symbols and
strings behave same (at least similar), and it had broken too many
programs. I understand your problem but it's not worth raising huge
compatibility issues.

 					matz.

Thank you for your feedback, Matz. I'd really never suspect such a
change could break so many programs...

I'm really curious what kind of programs rely on symbols behaving
differently from strings. But I guess you won't remember which programs
would be them, will you?

Updated by drbrain (Eric Hodel) about 11 years ago

On Feb 7, 2013, at 10:43, David MacMahon wrote:

I think the number of real world uses of something like:

{:a => 0, 'a' => 1}

is likely to be very small.

Every time you run gem you use a hash that contains both symbol and string keys used for different purposes.

In ~/.gemrc symbols are used for configuration options while strings are used for command default arguments.

Updated by Anonymous about 11 years ago

Hi,

In message "Re: [ruby-core:52017] Re: [ruby-trunk - Feature #7792] Make symbols and strings the same thing"
on Fri, 8 Feb 2013 09:26:38 +0900, Rodrigo Rosenfeld Rosas writes:

|Thank you for your feedback, Matz. I'd really never suspect such a
|change could break so many programs...
|
|I'm really curious what kind of programs rely on symbols behaving
|differently from strings. But I guess you won't remember which programs
|would be them, will you?

It was long long ago. But some programs distinguished symbols and
strings as hash keys, for example.

Symbols are taken from Lisp symbols, and they has been totally
different beast from strings. They are not nicer (and faster)
representation of strings. But as Ruby stepped forward its own way,
the difference between symbols and strings has been less recognized by
users.

						matz.

Updated by david_macmahon (David MacMahon) about 11 years ago

On Feb 7, 2013, at 5:15 PM, Eric Hodel wrote:

On Feb 7, 2013, at 10:43, David MacMahon wrote:

I think the number of real world uses of something like:

{:a => 0, 'a' => 1}

is likely to be very small.

Every time you run gem you use a hash that contains both symbol and string keys used for different purposes.

In ~/.gemrc symbols are used for configuration options while strings are used for command default arguments.

Well, that's one! In my defense, one is a very small number even if that one is of very large consequence. :-)

Just to play Devil's advocate, could that not be separated into two different hashes: one for config options and one for command default arguments? You have given semantic meaning to (the class of) the key. Isn't that akin to treating the primary key of a database record in a non-opaque manner?

Dave

Updated by spatulasnout (B Kelly) about 11 years ago

Rodrigo Rosenfeld Rosas wrote:

I'm really curious what kind of programs rely on symbols behaving
differently from strings.

One example:

Since symbols aren't garbage collected, my Ruby-based RPC system
is designed by default to convert symbols in incoming messages to
strings instead. (PacketBuf.preserve_symbols_on_extract = false)

This has led to sending certain internal messages with names
consisting of symbols instead of strings. These messages cannot
be spoofed by an external source, as there's no way for a remote
sender to produce messages locally containing symbols.

(If not for the garbage collection problem, I would have designed
the system to preserve externally generated symbols, and so would
have then required a different approach to distinguish internal
vs. external message names. So current design is arbitrary in
that sense, but nevertheless, the symbol/string distinction is
being put to use.)

Regards,

Bill

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Em 07-02-2013 23:57, Bill Kelly escreveu:

...but nevertheless, the symbol/string distinction is being put to use.)

Thanks for letting me know, Bill.

Updated by drbrain (Eric Hodel) about 11 years ago

On Feb 7, 2013, at 17:37, David MacMahon wrote:

Just to play Devil's advocate, could that not be separated into two different hashes: one for config options and one for command default arguments? You have given semantic meaning to (the class of) the key. Isn't that akin to treating the primary key of a database record in a non-opaque manner?

Two hashes would not be backwards compatible with previous versions of RubyGems. I'm hesitant to change the format of a configuration file and break downgrades between RubyGems versions.

It would be more confusing to new users. Now we can tell users "paste '…' into ~/.gemrc to do …" and it will work. They don't need to know anything about ruby or yaml to do this.

If matz decides that we should break backwards compatibility in this manner I will change my mind.

Updated by david_macmahon (David MacMahon) about 11 years ago

On Feb 8, 2013, at 4:50 PM, Eric Hodel wrote:

On Feb 7, 2013, at 17:37, David MacMahon wrote:

Just to play Devil's advocate, could that not be separated into two different hashes: one for config options and one for command default arguments? You have given semantic meaning to (the class of) the key. Isn't that akin to treating the primary key of a database record in a non-opaque manner?

Two hashes would not be backwards compatible with previous versions of RubyGems. I'm hesitant to change the format of a configuration file and break downgrades between RubyGems versions.

It would be more confusing to new users. Now we can tell users "paste '…' into ~/.gemrc to do …" and it will work. They don't need to know anything about ruby or yaml to do this.

If matz decides that we should break backwards compatibility in this manner I will change my mind.

I was not trying to induce change. I think my follow-on query would have been more appropriate for ruby-talk rather than ruby-core. I will (try to) keep my ruby-core postings more on topic.

Thanks for the explanation,
Dave

P.S. Sorry if this gets out more than once. I'm in the midst of an involuntary email "transition".

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Just another confusion caused by different behavior of strings vs symbols: https://github.com/blog/1440-today-s-email-incident

Updated by Student (Nathan Zook) about 11 years ago

The problem is not "confusion", it is that there is an apparent effort to rework core features in a massive framework in little pieces. These things don't fail in isolation--it is an error to attempt to fix them in isolation.

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

What I meant is that most people only use symbols as an identifier that could be a just a string as well. If you look at the code semanthic you'll notice that they don't really care about symbol vs strings differences. They just want an identifier but then they have to be concerned about their differences all over the code even when they don't want this because different pieces will have different opinions about if symbols or strings should be used for that case. Sometimes they only happen to use symbols because of the convenience of the new hash better looking syntax. I'm just trying to convince Matz to evaluate once again the possibility of making symbols behave like string.freeze! or simply like regular strings even if it would break backward compatibility... I still believe all the confusion created by their differences will keep happening and breaking compatibility for good in this case makes sense to me. Please reconsider this ticket...

Updated by rosenfeld (Rodrigo Rosenfeld Rosas) about 11 years ago

Also, as a nice side effect we wouldn't even have to worry about garbage collecting symbols to avoid DoS as they would be automatically collected just like strings. If MRI wants to keep the original notion of symbols as an internal implementation detail, that's fine, but this shouldn't leak to end users...

Updated by Student (Nathan Zook) about 11 years ago

Actually, rails trusts symbols at times in ways that it does not trust strings. This is the source of a recent security issue or two, as it was assumed that symbols creation would not be directed from outside data.

Better would have been to have run in safe mode 1, or to have watched the tainted property directly.

Without studying this most recent issue in detail, it looks like the rails devs have opted for rapid, "mostly right" instead of deeply-analysed, thoroughly-vetted solutions. Given the nature of some of these issues, that was almost certainly correct as an initial response.

Updated by alexeymuranov (Alexey Muranov) about 11 years ago

Student (Nathan Zook) wrote:

Actually, rails trusts symbols at times in ways that it does not trust strings. This is the source of a recent security issue or two, as it was assumed that symbols creation would not be directed from outside data.

In Rails, the Hash#symbolize_keys! method looks somewhat unacceptable to me.

Updated by Student (Nathan Zook) almost 11 years ago

Yep. If to_sym threw or nilled on previously unknown symbols, or at least on tainted data, this would be okay. Otherwise, it is a serious security problem.

Updated by wardrop (Tom Wardrop) almost 11 years ago

=begin
= To Summarise...

== Symbol Differences

=== Immutable
A symbol is an alias for a numeric ID. The symbol representation of that numeric ID can never change.

=== Not garbage collected
An extension of the previous point. Garbage collecting symbol's would break the fact that symbol's are an alias for a specific numeric ID.

=== Semantically different
The type difference between a String and a Symbol can be used in program logic.

== The Problem
((And yes, there is clearly a problem! You can argue the solution.))

=== Inconsistance use
It's virtually impossible to reliably determine when to use a Symbol or a String. Today, Symbol might be the logical choice for scenario X, but tomorrow you're looking up those Symbols with user-entered data, or reading them in from a file, etc. Programs change, requirements change, everything changes.

=== Library incompatibility
As a consequence of the previous point, combined with different levels of programmer experience and plain personal opinion, libraries will always contain incompatibility.

=== Unignorable (((It's a real word, I swear)))
As a consequence of the previous two, even if you only use one or the other, unless you want to avoid using 3rd party code, stdlib, or even ruby-core, there's no way to ignore the fact that Symbol's are different and completely incompatible with String's. Pita.

== The Solution
Or more to the point, what criteria does a solution need to meet. I don't believe any solution can resolve all the problems, while maintaing full backwards compatibility. Therefore, I see there being a short-term solution, i.e. minor version release of 2.x, and a long term solution, i.e 3.0.

I think the Ruby project needs to maintain a list of desired backwards-incompatible changes, so these can be rolled up into one big 3.0 release. Ruby can't be backwards compatible forever, otherwise it'll slowly fade into insignificance... hey Microsoft.

=== Short Term
I don't think there is one. HashWithIndifferentAccess will not solve the problem with hash keys. You'll just go from coercing Symbols to String and vice versa, to coercing Hash to HashWithIndifferentAccess, and back again, and there's no way we can make the differences ignorable to those libraries that don't care, while still satisfying those libraries that do care. Anything we change will probably only help to dig us all a deeper hole.

The short term solution is to accept there's a problem, and that anything backwards-compatible isn't worth changing.

=== Long Term
I propose the remainder of this thread should focus on a long term solution that solves ALL the problems without worrying about full backwards-compatibility, and for everyone to accept that there is indeed a problem with the existence and/or current implementation of Strings and Symbols. Getting rid of Symbols, making Symbols a sub-class of String, whatever. Let's talk about that now, please. It's clear that even Matz is unsatisfied with the current situation.

=end

Updated by phluid61 (Matthew Kerwin) almost 11 years ago

=begin
In that case, I suggest that the long term solution is education.

By the way: "((A symbol is an alias for a numeric ID.))" Not entirely true (otherwise we could argue for Symbol#to_i or even Symbol#to_int) A symbol represents an immutable concept, analogous to a number (i.e. 3 is always 3, even outside of Ruby; so :to_s is always :to_s). The numeric ID is an implementation detail.

Also, it's spelled "Summarise" with an 'a' (or "Summarize" in America.) ;)
=end

Updated by wardrop (Tom Wardrop) almost 11 years ago

No that's spelt correctly. I sadly couldn't embed any pictures of palm trees and coconuts ;). Fixed, thanks.

Updated by spatulasnout (B Kelly) almost 11 years ago

wardrop (Tom Wardrop) wrote:

=== Immutable
A symbol is an alias for a numeric ID. The symbol representation of that
numeric ID can never change.

As Matthew Kerwin noted, the numeric ID is an implementation detail; and
indeed, the IDs already do change:

$ ruby -e "p [:foo.object_id, :bar.object_id]"
[143928, 144008]

$ ruby -e "p [:bar.object_id, :foo.object_id]"
[143928, 144008]

=== Not garbage collected
An extension of the previous point. Garbage collecting symbol's would break
the fact that symbol's are an alias for a specific numeric ID.

Presumably this could only affect programs which retain the object_id
of a symbol past the point where the symbol itself is no longer
referenced by the program, and which then expect to be able to compare
that object_id with the object_id of the equivalent symbol after it
has been GC'd and re-interned.

Assuming the rest of the GC-related problems have been solved, such as
marking any symbols interned by ruby extensions as non-GC-able, then
we should only need wonder how much pure ruby code exists that, during
the lifetime of a given ruby interpreter process, is storing and
comparing symbol object_ids rather than the symbols themselves.

If much such code indeed exists in production environments, then
presumably symbol GC could be introduced as a VM startup option,
eventually becoming the default after the legacy behavior has gone
through a deprecation phase.

It's not evident to me that a solution to the symbol GC issue needs
to be bundled with solution(s) for symbol/string equivalence issues,
however.

Regards,

Bill

Updated by wardrop (Tom Wardrop) almost 11 years ago

@spatulasnout (B Kelly) Yes on your first point. One should always compare :symbol with :symbol rather than by their internal ID, so I'd say that any code using the internal ID's deserves to break. I imagine it would be hardly any, other than the interpreter itself.

So on that note, I think you're right. Garbage collecting symbol's could be implemented while maintaining backwards compatibility. That may address the technical implications such as the potential for DoS attacks. As you said, this could safely be implemented as a startup option, or potentially even a runtime option.

Updated by Ajedi32 (Andrew M) over 9 years ago

So now that Symbol GC is implemented (#9634), does that change anything about the viability of this proposal?

Updated by eloyesp (Eloy Esp) over 9 years ago

I think the proposal is goood, if the change is better to new programmers it means it is better for all programmers, (because eventually the old experienced ruby programmer will die) and we will be adapting the people to the code instead of the other way around (that is not a good practice).

As I understand, the proposed change can be read as:

  • Currently there is an exception in the language for strings as hash keys (they are duped and frozen).
  • This behaviour is unexpected for new rubysts and for not so new (I learned that yesterday, and I worked with ruby for 4+ years).
  • Using strings as symbols for hash keys is a headache for many programmers. (there are a lot of libraries to make this easier).
  • There is a lot of libraries written with the traditional way in mind. Thus all the gems that stores in the same hash strings and symbols could break if we change the exception.

Given the change will break things, but will make things better, why not plan it for ruby 3.0?

The only use-case that will break things is:

hash = {}
hash['key'] = value
hash[:key] = other_value

We can add a warning when you do something like this and make the change a year after that.

Updated by subtileos (Daniel Ferreira) about 7 years ago

Hi Bill

One example:

Since symbols aren't garbage collected, my Ruby-based RPC system
is designed by default to convert symbols in incoming messages to
strings instead. (PacketBuf.preserve_symbols_on_extract = false)

This has led to sending certain internal messages with names
consisting of symbols instead of strings. These messages cannot
be spoofed by an external source, as there's no way for a remote
sender to produce messages locally containing symbols.

(If not for the garbage collection problem, I would have designed
the system to preserve externally generated symbols, and so would
have then required a different approach to distinguish internal
vs. external message names. So current design is arbitrary in
that sense, but nevertheless, the symbol/string distinction is
being put to use.)

With symbols being garbage collected now is this functionality broken or are you still relying on symbols over strings for your solution?

Thanks,

Daniel

Updated by spatulasnout (B Kelly) about 7 years ago

Hi Daniel,

Daniel Ferreira wrote:

(If not for the garbage collection problem, I would have designed
the system to preserve externally generated symbols, and so would
have then required a different approach to distinguish internal
vs. external message names. So current design is arbitrary in
that sense, but nevertheless, the symbol/string distinction is
being put to use.)

With symbols being garbage collected now is this functionality broken
or are you still relying on symbols over strings for your solution?

I've continued using the aforementioned RPC system without enabling
the setting that would allow it to recreate symbols on the receiving
end.

So it continues to work as it did before (if the remote sends symbols,
they appear only as strings locally.)

I will say however that on the balance, in the time I've been using
this system, it can feel sub-optimal to have to keep track of which
parameters are symbols and which are strings. I'd say my primary
motivation for using producing symbols on the sending end is Ruby's
convenient hash syntax:

result = some.rpc_method(foo:123, bar:"baz")

I find that i like the syntax enough, that I'll put up with having
to think about when those symbols will have been turned into strings
down the line.

But it would be nice to not have to keep track of that detail.

Regards,

Bill

Updated by subtileos (Daniel Ferreira) about 7 years ago

result = some.rpc_method(foo:123, bar:"baz")

I find that i like the syntax enough, that I'll put up with having
to think about when those symbols will have been turned into strings
down the line.

I believe it is time now with the advent of Ruby 3 to revisit this feature request.
Although maybe not quite like it is stated right now.

What are your feelings about it?

Updated by sheerun (Adam Stankiewicz) over 6 years ago

@shyouhei (Shyouhei Urabe) @matz (Yukihiro Matsumoto) Could this be revisited as ruby now has frozen strings? They behave pretty much the same as symbols.

Here's an article worth read: http://blog.arkency.com/could-we-drop-symbols-from-ruby/

Updated by shevegen (Robert A. Heiler) over 6 years ago

I found the old issue here too from the blog article. :)

However had, I do not think that frozen strings would be the same (semantic-wise) as Symbols so I am
not sure that this would lead to a reconsideration of the String/Symbol distinction.

I personally am fine with the way how Symbols behave as-is; I do also understand that other people
think of the distinction somewhat annoying - the best example may be HashWithIndifferentAccess :D

I mean, that is so long to write out, I wonder how people can use it over "Hash"... but I understand
the problem situation. People want to squeeze out whatever is faster, and also not have to think "when
to use strings and when to use symbols" - it's like an obsession and somewhat understandable the larger
(and perhaps slower) a codebase becomes + amount of users of said codebase. The mental "cognitive load"
will remain as long as ruby has a string-versus-symbols distinction, but I myself have come to like
symbols. :cat and :dog versus 'cats' and 'dogs'.

My only minor concern with the distinction was whether I would be able to treat Symbols as String-like
for some string-like operations, such as via :foobar[1,2] - but that is already possible. Walks
like a duck, quacks like a duck, is a symbolized duck. Actually, an even more radical solution would
be to get rid of Symbols altogether, since that would simplify the usage instantly. Are frozen strings
and symbols really as fast? I have no benchmarks but I think that symbols may still be faster, since
they may do more - but since I have no benchmark, this is just idle speculation. I mention this
because the old proposal here wrote "make symbols and strings the same thing" but you could also
have a proposal that would "remove symbols" ... :P

Matz explained the origin of Symbols back then here:

http://blade.nagaokaut.ac.jp/cgi-bin/scat.rb/ruby/ruby-core/52019

I think Symbols became more popular after the Dave-book stating that they are much faster
in Hashes than string-object, so since people like speed, they'd use Symbols a lot. I
like symbols a lot too, but mostly because ... it's less to type! I can also enable some
specific behaviour depending on input arguments given to methods. This may be a very
akward usage example but I like it.

Perhaps you may have a chance to see the situation be revisited for ruby 3.x but I myself
think that the Symbols will stay as they are for other reasons, including inertia. Another
factor to be considered is how ruby "views" the its world - and I think that the original
idea never was to expose symbols as-is and instead may have users focus only on strings
primarily.

Btw here is a short discussion at stackoverflow about symbols in lisp:

https://stackoverflow.com/questions/8846628/what-exactly-is-a-symbol-in-lisp-scheme

What I gather there is that the primary use case is indeed due to speed (or "light-weight
execution") compared to alternatives.

Updated by dsferreira (Daniel Ferreira) over 6 years ago

If we are considering breaking backwards compatibility in ruby 3.0 with removal of API like IO.read which is all over let’s fix this symbol vs string issue once and for all shall we?
Ruby programmers will be much happier.
Why don’t we start a new issue maybe in common ruby to come up with a broad solution?

Updated by dsferreira (Daniel Ferreira) over 6 years ago

If we do want to keep symbols as part of ruby API what about considering a new syntax for symbols and leave the current symbol syntax to be identified as string literals or just strings?

Updated by normalperson (Eric Wong) over 6 years ago

wrote:

If we are considering breaking backwards compatibility in ruby
3.0 with removal of API like IO.read which is all over let’s

Removal of IO.read? I hope you're not misinterpreting my [ruby-core:84461]
response to https://bugs.ruby-lang.org/issues/14239
I only want to warn on the IO.read("| command") case,
not remove the whole thing.

Anyways, I do not have much opinion on symbols/strings;
but I hate breaking compatibility.

Updated by dsferreira (Daniel Ferreira) over 6 years ago

It seems that what is in question to be removed in ruby 3.0 is the pipe feature associated to Kernel#open and related API not the all method.

Still breaking backwards compatibility though.

Sorry for misreading it and thanks Eric for pointing me out the error.

https://bugs.ruby-lang.org/issues/14239

Updated by normalperson (Eric Wong) over 6 years ago

Daniel Ferreira wrote:

Why isn’t your reply on the issue tracker?

It should be, I see it. Might take some minutes for mail to propagate.

Updated by ko1 (Koichi Sasada) over 6 years ago

It seems that what is in question to be removed in ruby 3.0 is the pipe feature associated to Kernel#open and related API not the all method.

Still breaking backwards compatibility though.

Important point is, we can know the breaking compatibility with exceptions. If we remove some method or feature (like IO.open('|...') we (the application programmer) can understand it was obsolete.

However, this case, program breaks silently and difficult to debug.

Personally I like this idea and if I have a time machine, I wanted to suggest this idea at 199x.
If we need to introduce breaking compatibility, we need to make transition path.

Updated by dsferreira (Daniel Ferreira) over 6 years ago

If we need to breaking compatibility, we need to make transition path

Thanks Koichi.
I totally agree with you and Eric about worries on breaking backwards compatibility.
I was wondering if it wouldn’t be possible to use some kind of transitional flag like:

# frozen_string_literal: true
Or
--enable-frozen-string-literal

Updated by ko1 (Koichi Sasada) over 6 years ago

I was wondering if it wouldn’t be possible to use some kind of transitional flag like:

Please consider transition path for users who are using symbol and string difference like:

key = ...
...
when key
case String
  ...
case Symbol
  ...
end

How to find out such programs?

frozen-strign-literal, we can know issues by seeing FrozenError (and also debug_frozen_string_literal is supported to find out the problem).

If you (or someone) find out any good transition path, we think we can consider again.

Updated by dsferreira (Daniel Ferreira) over 6 years ago

Hi Koichi.

I’ve been thinking about the challenge you put me.
What if we add as a transition path the method:
String#old_symbol.

Then we could use it in the Symbol::===(var)
and we would warn if var.is_a?(String) && var.old_symbol == true

What do you think?
Something missing?

Updated by jeremyevans0 (Jeremy Evans) over 6 years ago

I'm 100% against merging Symbols and Strings. Symbols and Strings have always represented different concepts. Symbols are a simple way to get a unique named identifier, and Strings represent arbitrary data/text. In some cases, you need to convert from one or the other, generating an identifier from the content of data/text (String->Symbol), or treating the name of the identifier as data/text (Symbol->String). But Strings and Symbols represent different concepts and should be treated differently.

Merging Symbols and Strings would break libraries that do treat them differently in ways that would be almost impossible to fix without making APIs very ugly (and obviously breaking backwards compatibility).

The main argument I've heard for combining Strings and Symbols is to save some characters when typing, as shown in the initial post for this issue. I don't consider this a major problem, but if other people really want to save those characters, I think it would be better to add a terser form of string literals, as opposed to completely breaking symbols.

Updated by dsferreira (Daniel Ferreira) over 6 years ago

But Strings and Symbols represent different concepts and should be treated differently.

Jeremy I agree with you on this and that is one of the reasons, possibly the biggest reason, why I think it will make a huge difference if we stop using symbols as if they were plain simple strings like I see scattered all over the place.

Let’s put symbols where they belong shall we?

I ask again: Shall we open a new issue under common ruby with a broader purpose?

The reality is that converting current symbol’s syntax to be used as yet another string’s syntax will keep intact the great majority of the code currently in production.

From this discussions we don’t see many examples of proper use of symbols as such.

I believe we can fix this and keep Symbols and Strings as part of ruby API in a much better way.

A new syntax for symbols is the first thing that comes to my mind or a plain Symbol.new instantiation.

Updated by jeremyevans0 (Jeremy Evans) over 6 years ago

dsferreira (Daniel Ferreira) wrote:

But Strings and Symbols represent different concepts and should be treated differently.

Jeremy I agree with you on this and that is one of the reasons, possibly the biggest reason, why I think it will make a huge difference if we stop using symbols as if they were plain simple strings like I see scattered all over the place.

I think it's far more common to use symbols correctly than incorrectly. The only time I see confusion is when using things like HashWithIndifferentAccess. That more than anything else has probably contributed to some ruby programmers not understanding what the difference is.

I ask again: Shall we open a new issue under common ruby with a broader purpose?

You are free to open an issue. However, considering this has already been rejected, I don't see the point.

The reality is that converting current symbol’s syntax to be used as yet another string’s syntax will keep intact the great majority of the code currently in production.

"Extraordinary claims require extraordinary evidence"

From this discussions we don’t see many examples of proper use of symbols as such.

I'm not sure if this discussion is reflective of symbol usage in most ruby code. In most cases I see symbols used correctly, as named identifiers. In most cases I see strings used correctly, as text/data. There are a few cases where I see them used incorrectly, mostly at the borders of programs, where input is always provided as strings. However, that's a security sensitive area and care should be taken before converting any strings to symbols.

I believe we can fix this and keep Symbols and Strings as part of ruby API in a much better way.

A new syntax for symbols is the first thing that comes to my mind or a plain Symbol.new instantiation.

This does not make sense to me. If you introduce a new syntax for symbols, most people won't switch to it unless they don't care about backwards compatibility at all. If by some miracle the community does switch to it, you'll end up the same situation we currently have eventually.

You've already recognized that strings and symbols are different concepts and have different purposes. Therefore the only reason make the current symbol syntax represent strings if you want a terser syntax for strings. In which case instead of breaking backwards compatibility, it would make more sense to introduce a new separate terser syntax for strings.

Updated by dsferreira (Daniel Ferreira) over 6 years ago

I think it's far more common to use symbols correctly than incorrectly. The only time I see confusion is when using things like HashWithIndifferentAccess.

Let me give you just this example to show you how things are out of control:

alias_method :foo, :bar
alias_method "foo", "bar"

Both usages are allowed.

This kind of interface is present in many places.
You have it in ruby core and you have it even more in gems and even more in applications.
And there will be no way you can change that line of thought scattered throughout the community.
We are using strings and symbols interchangeably and this is an evidence.

Updated by Ajedi32 (Andrew M) over 6 years ago

Leaving aside the issue of semantics for a moment, can anyone provide a real, concrete example of a situation where symbols could not trivially be replaced with strings without negatively impacting the functionality of the surrounding code?

I'm having trouble thinking of such a scenario that's not obviously contrived.

Updated by dsferreira (Daniel Ferreira) over 6 years ago

Leaving aside the issue of semantics for a moment, can anyone provide a real, concrete example of a situation where symbols could not trivially be replaced with strings without negatively impacting the functionality of the surrounding code?

Koichi’s challenge for a transition path is the example we have to support so far:

key = ...
...
when key
case String
  ...
case Symbol
  ...
end
Actions

Also available in: Atom PDF

Like0
Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0