问题
I have 50 sidekiq threads crawling the web, and a few weeks ago the threads started hanging after about 20 minutes of running. When I do a backtrace dump, most of the threads are stuck on net/http initialize:
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/net/http.rb:879:in `initialize'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/net/http.rb:879:in `open'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/net/http.rb:879:in `block in connect'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:76:in `timeout'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/net/http.rb:878:in `connect'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/net/http.rb:863:in `do_start'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/net/http.rb:858:in `start'
/app/vendor/bundle/ruby/2.1.0/gems/net-http-persistent-2.9.4/lib/net/http/persistent.rb:700:in `start'
/app/vendor/bundle/ruby/2.1.0/gems/net-http-persistent-2.9.4/lib/net/http/persistent.rb:631:in `connection_for'
/app/vendor/bundle/ruby/2.1.0/gems/net-http-persistent-2.9.4/lib/net/http/persistent.rb:994:in `request'
/app/vendor/bundle/ruby/2.1.0/gems/mechanize-2.7.2/lib/mechanize/http/agent.rb:257:in `fetch'
/app/vendor/bundle/ruby/2.1.0/gems/mechanize-2.7.2/lib/mechanize/http/agent.rb:974:in `response_redirect'
/app/vendor/bundle/ruby/2.1.0/gems/mechanize-2.7.2/lib/mechanize/http/agent.rb:298:in `fetch'
/app/vendor/bundle/ruby/2.1.0/gems/mechanize-2.7.2/lib/mechanize.rb:432:in `get'
/app/app/workers/crawl_page.rb:24:in `block in perform'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:91:in `block in timeout'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:35:in `block in catch'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:35:in `catch'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:35:in `catch'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:106:in `timeout'
I didn't think sidekiq would get stuck on net/http because I've wrapped the entire call in a timeout: Timeout::timeout(APP_CONFIG['crawl_page_timeout']) { @page = agent.get(url) }
...but then I started reading some old posts about how ruby's Timeout is not thread safe: http://blog.headius.com/2008/02/rubys-threadraise-threadkill-timeoutrb.html
Is ruby's Timeout still not thread safe?
I know a lot of people write crawlers in Ruby. If Timeout isn't thread-safe, how are people writing crawlers handling the issue of net/http getting stuck?
Update:
I've switched to HTTPClient (which specifically says its thread safe) to replace mechanize. We appear to still be getting stuck on initializing a thread. Again, this could be due to ruby'ss Timeout not working properly, or it could be a sidekiq issue. Here's the stacktrace from the most recent hung sidekiq threads:
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient/session.rb:805:in `initialize'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient/session.rb:805:in `new'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient/session.rb:805:in `create_socket'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient/session.rb:752:in `block in connect'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:91:in `block in timeout'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:101:in `call'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:101:in `timeout'
/app/vendor/ruby-2.1.2/lib/ruby/2.1.0/timeout.rb:127:in `timeout'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient/session.rb:751:in `connect'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient/session.rb:609:in `query'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient/session.rb:164:in `query'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient.rb:1087:in `do_get_block'
/app/vendor/bundle/ruby/2.1.0/gems/newrelic_rpm-3.9.2.239/lib/new_relic/agent/instrumentation/httpclient.rb:34:in `block in do_get_block_with_newrelic'
/app/vendor/bundle/ruby/2.1.0/gems/newrelic_rpm-3.9.2.239/lib/new_relic/agent/cross_app_tracing.rb:43:in `tl_trace_http_request'
/app/vendor/bundle/ruby/2.1.0/gems/newrelic_rpm-3.9.2.239/lib/new_relic/agent/instrumentation/httpclient.rb:33:in `do_get_block_with_newrelic'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient.rb:891:in `block in do_request'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient.rb:985:in `protect_keep_alive_disconnected'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient.rb:890:in `do_request'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient.rb:963:in `follow_redirect'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient.rb:776:in `request'
/app/vendor/bundle/ruby/2.1.0/gems/httpclient-2.4.0/lib/httpclient.rb:677:in `get'
/app/app/ohm_models/queued_page.rb:20:in `run_crawl'
回答1:
Correct, it is still not safe to use Timeout in Ruby code, unless you know exactly what is happening within that block (which includes what any C code might be doing). I have personally observed catastrophic things happening in connection pools because of this.
You may be able to get away with rescuing errors and retrying, but if you're unlucky your process might get wedged and require a restart.
If you fork to create new processes, you can kill those safely if they run long (or use timeout(1) because they don't have any way to corrupt your parent process.
I know a lot of people write crawlers in Ruby. If Timeout isn't thread-safe, how are people writing crawlers handling the issue of net/http getting stuck?
Do you have a specific example of one that works?
来源：https://stackoverflow.com/questions/25803089/is-ruby-2-1-2-timeout-still-not-thread-safe