Request thread safety issue

[paul]

We're getting weird connection issues in our exception tracker:

HTTP::ConnectionError: error reading from socket: stream closed in another thread
 per_operation.rb  79 wait_readable(...)
[GEM_ROOT]/gems/http-4.1.1/lib/http/timeout/per_operation.rb:79:in `wait_readable'
Caused by:
 IOError: stream closed in another thread
 per_operation.rb  79 wait_readable(...)
[GEM_ROOT]/gems/http-4.1.1/lib/http/timeout/per_operation.rb:79:in `wait_readable'

The logs for that single job don't show anything interesting, but if you look at the logs for all the workers in that process:

[2019-07-12T15:40:15.589244 #4]  INFO -- : [03e56584-c7a1-45ee-9ca4-c86334cd7e75] > POST https://onesignal.com/api/v1/notifications 
[2019-07-12T15:40:15.886281 #4]  INFO -- : [1bf0df80-9dde-4724-a377-5ed9271284de] > POST https://onesignal.com/api/v1/notifications 
[2019-07-12T15:40:15.887557 #4]  INFO -- : [03e56584-c7a1-45ee-9ca4-c86334cd7e75] HTTP::ConnectionError: error reading from socket: stream closed in another thread 
[2019-07-12T15:40:15.889464 #4] ERROR -- : [03e56584-c7a1-45ee-9ca4-c86334cd7e75] Error performing WisperActiveJobBroadcaster::Job (Job ID: 03e56584-c7a1-45ee-9ca4-c86334cd7e75) from Shoryuken(tesseract-production-background) in 691.77ms: HTTP::ConnectionError (error reading from socket: stream closed in another thread): 
[2019-07-12T15:40:16.268402 #4]  INFO -- : [1bf0df80-9dde-4724-a377-5ed9271284de] < 200 OK 

Job 03e5 started posting to the endpoint, and 300ms later another thread (job 1bf0) posted to the same endpoint with a different body. This somehow interrupted the first request in progress, raising the IOError in the first 1.5ms after the 2nd request started. The 2nd request then finished normally.

Here's how we're making that request:

# global default HTTP client
  MyAppHTTP =
    ::HTTP
    .timeout(connect: 2.seconds, read: 10.seconds, write: 10.seconds)
    .use(logging: { logger: Rails.logger })

# class-specific HTTP client

  def http
    @http ||= MyAppHTTP.
              headers(headers).
              timeout(write: 30, connect: 10, read: 40).
              auth("Basic \"#{api_key}\"")
  end

# called like this:

      http.post(url("/notifications"), json: payload.to_json(:web)).status.success?

We're using the shoryuken gem as our job runner, which uses concurrent-ruby for the worker thread pool. I'm not clear how making a request in one worker thread could interfere with another. The wiki says "Thread safety comes into play when you Make an HTTP request", but doesn't offer any details, or instruction how to mitigate it.

This happens more often than I would have expected, too:

[paul]

Reduced to a simple test case:

require "thread"
require "http"

@http = HTTP.headers("X-Test" => "1")

5.times.map do
  Thread.new { r = @http.post("https://httpbin.org/post"); puts r.status}
end.map(&:join)

This breaks in a similar way:

/home/rando/.rubies/ruby-2.6.3/lib/ruby/2.6.0/openssl/buffering.rb:125:in `sysread': stream closed in another thread (IOError)
	8: from test.rb:11:in `block (2 levels) in <main>'
	7: from /home/rando/.gem/ruby/2.6.3/gems/http-4.1.1/lib/http/chainable.rb:27:in `post'
	6: from /home/rando/.gem/ruby/2.6.3/gems/http-4.1.1/lib/http/client.rb:31:in `request'
	5: from /home/rando/.gem/ruby/2.6.3/gems/http-4.1.1/lib/http/client.rb:75:in `perform'
	4: from /home/rando/.gem/ruby/2.6.3/gems/http-4.1.1/lib/http/connection.rb:103:in `read_headers!'
	3: from /home/rando/.gem/ruby/2.6.3/gems/http-4.1.1/lib/http/connection.rb:212:in `read_more'
	2: from /home/rando/.gem/ruby/2.6.3/gems/http-4.1.1/lib/http/timeout/null.rb:45:in `readpartial'
	1: from /home/rando/.rubies/ruby-2.6.3/lib/ruby/2.6.0/openssl/buffering.rb:125:in `readpartial'
/home/rando/.rubies/ruby-2.6.3/lib/ruby/2.6.0/openssl/buffering.rb:125:in `sysread': error reading from socket: stream closed in another thread (HTTP::ConnectionError)

Full console output

While modifying the thread block to use the HTTP constant directly works just fine.

5.times.map do
  Thread.new { r = HTTP.post("https://httpbin.org/post"); puts r.status}
end.map(&:join)

I would have expected to be able to configure the HTTP client once and be able to re-use it amongst threads, the same way calling the HTTP class methods work. Do I need to set up a concurrent-ruby thread pool of these configured clients?

[tarcieri]

Possibly related to #371

[tarcieri]

To try to answer this:

I would have expected to be able to configure the HTTP client once and be able to re-use it amongst threads, the same way calling the HTTP class methods work. Do I need to set up a concurrent-ruby thread pool of these configured clients?

The original goal of the HTTP::Options API builder was for it to be immutable and thread safe, and have additional chained options make copies of the original state rather than mutating. Re #371, it seems like there are places this isn't happening and the state is being shared? I'm not entirely sure.

HTTP::Client isn't thread safe, and stores multiple different kinds of mutable state. Persistent connections are likewise not thread safe, and if you'd like to use them in a multithreaded context I'd suggest using something like the connection_pool gem.

[paul]

@tarcieri So do you think the HTTP gem should allow the simple test case in my comment to work or not? (The one where I re-use the same @http ivar in multiple threads to make simultaneous requests, no persistent connections involved.)

I'll start using the ConnectionPool for this, I just found the behavior surprising.

[tarcieri]

@paul if I’m reading your example right, @http should be an instance of HTTP::Options, and therefore your example should work. I’m curious if you think #371 is related/the same problem

[paul]

It's entirely possible. Maybe Ruby in 2015 couldn't detect this problem so just returned an empty string, and modern Ruby IO notices somehow and is able to raise the IOError.

[tarcieri]

Re-reviewing #371 again, I think I was thinking of a different issue...

[tarcieri]

@paul so the real question is...

HTTP::Options is (or was at least originally intended to be) thread-safe by virtue of being immutable.

Are you encountering thread-safety problems with it? (i.e. is it your @http)

[ixti]

FWIW HTTP::Client itself isn't thread-safe:

http/lib/http/client.rb

Lines 71 to 76 in 6240672

	@connection ||= HTTP::Connection.new(req, options)
	unless @connection.failed_proxy_connect?
	@connection.send_request(req)
	@connection.read_headers!
	end

[ixti]

In order to make thread-safe concurrent calls with same HTTP::Client instance, you will need to use mutex, something like this:

@mutex = Mutex.new

# ...

res = @mutex.synchronize { http.get(...).flush }

In other words you need to guarantee that there's no concurrent request between current request and Response#flush call.

[tarcieri]

Yes, so HTTP::Options was originally intended to be immutable/thread safe, and HTTP::Client is not since it owns the (mutable) connection state, however the idea was you could use HTTP::Options as a request builder and then pass the result to different threads and instantiate thread-specific HTTP::Client instances there by making requests.

If I'm understanding correctly though, @paul is trying to do the first thing, which is share HTTP::Options between threads, and then instantiate new clients in each thread... and having thread safety issues with that.

@paul can you confirm that's the case?

[ixti]

Mmm, but the snippet shared above is absolutely not re-use of options, but re-use of client:

@http = HTTP.headers("X-Test" => "1")

5.times.map do
  Thread.new { r = @http.post("https://httpbin.org/post"); puts r.status}
end.map(&:join)

[tarcieri]

@ixti oh, now I see...

https://github.com/httprb/http/blob/master/lib/http/chainable.rb#L250

I guess that changed way back in #7

[tarcieri]

@paul this seems like an unfortunate complication of there not being a proper "builder" type for the client, allowing options which don't require the creation of e.g. sockets to be performed the way you want.

It seems like retrofitting an API like that is possible, by splitting up the the HTTP::Chainable#branch method to return HTTP::Options in cases where it doesn't need to eagerly instantiate a client.

[ixti]

The only thing that might not work correctly with that idea is keep-alive (persistent) clients...