How to avoid blocking Nginx with long request
jefftk at google.com
Wed Apr 10 13:56:40 UTC 2013
Why is your module taking a long time? Is it doing heavy computation
or is it blocked on IO? If it's blocking IO, can you rewrite it to
use asynchronous IO and never block?
Another option would be to put your code in a separate process and
reverse proxy to it.
Or you could be crazy and do what ngx_pagespeed does: use threads.
This is very tricky to get right.
On Wed, Apr 10, 2013 at 9:40 AM, Maxim Dounin <mdounin at mdounin.ru> wrote:
> On Wed, Apr 10, 2013 at 01:26:15PM +0000, MAGNIEN, Thierry wrote:
>> I'm writing an Nginx module that uses information stored in
>> memory to redirect requests to other servers. Basically when a
>> GET requests arrives, it makes some checks and decides to which
>> Location the requests shall be redirected. In order to have
>> Nginx update the information it holds in memory, I send him a
>> specific POST request to trigger it.
>> However, reloading information takes quite a lot of time and I
>> have some questions related with this:
>> - while the POST request is handled in my module, the worker
>> that took the request is blocked until it has finished
>> processing, but if GET requests come in, are they handled by
>> other workers or can I have some GET requests getting blocked ?
> If a worker process is blocked, it can't accept new connections.
> But connections already accepted, e.g. just before the POST
> request in question, are bound to the worker and will be blocked
> till you return to the event loop.
>> - if I want the processing not to block, can I use an event
>> timer, in order to release the worker quickly and have the
>> processing take place "in background" ? Or will it block a
>> worker anyway ?
> There is no such thing as "in background". As nginx is event
> driven, everything happens in event handlers, and timers are
> events as well. That is, just running a task in a timer handler
> will equally block the worker process.
> What might help is splitting the task into smaller ones and
> returning to the event loop inbetween, to let other requests time
> to be served. This may be done e.g. via 1ms timer.
> Alternatively, you may consider doing a reload with standard nginx
> configuration reload mechanism. This way all blocking operations
> are done in master process, then new workers with updated
> configuration are spawned to handle requests.
> Maxim Dounin
> nginx-devel mailing list
> nginx-devel at nginx.org
More information about the nginx-devel