Server congestion

[seanlinsley]

Lately we've been experiencing quite a few incidents where the server becomes so congested that it becomes completely unresponsive.

Today at 1 PM Central Time: (IRC)

Today at 5 PM Central Time: (IRC)

In both cases we had to restart the server for it to get back to normal. We need to get to the bottom of this. What's causing the slowdown? Is there any insight to be gained from the log of pages accessed?

We should have a system to automatically alert us when the site is having such troubles.

[chadwhitacre]

We should have a system to automatically alert us when the site is having such troubles.

Pretty sure that's Pagerduty (#2072).

[seanlinsley]

Yeah, @patcon mentioned the details shortly after in IRC

[chadwhitacre]

I've marked this DevX ★.

[patcon]

"Site unavailable" notifications with pagerduty are simple, as we're converting UptimeRobot alert emails into pagerduty notifications. Just need to work with @whit537 and @zwn to get them properly set up and get on-call schedules spread fairly. (ie. #2072)

As for more advanced alerting on certain conditions or logs messages, that requires new tooling (graylog2? some heroku add-on for log monitoring? newrelic? etc.). We would still route any other services' monitoring through the consistent pagerduty alerting system. Let's make this issue about that.

I've added a Todo section to the OP

[zbynekwinkler]

Not writing to the database on each and every request for signed in users (even for static files, even for 304) could help us move further away from the congestion issue (see #2041). The query to update session_expires is 9th most often called query, while the first two most often called queries are BEGIN and COMMIT and 8th is SET client_encoding TO 'UTF8' so it is more like 6th really (when I count only the meaningful queries).

[seanlinsley]

@zwn I assume you queried the production database to find the most common queries? Could you list them out somewhere?

[seanlinsley]

@patcon assigned this ticket to himself, taking the intention here to set up proper alerting. I think that's nicely encapsulated in #2072 already, so I'm un-assigning @patcon from this ticket and removing it from working, as no one has started really working on the core issue here. (IRC)

[grampajoe]

I strongly recommend NewRelic for diagnosing these issues. The free NewRelic Standard you get with Heroku gives you things like detailed error reports, alerting, event (downtime, high error rate, etc.) reports, transaction tracing, etc..

[chadwhitacre]

IRC

[Changaco]

Closing. The specific issue of becoming completely unresponsive when busy threads are maxed out has been fixed by #2384.