On Oct 18, 2009, at 11:31 AM, Benoit Chesneau wrote:

On Sun, Oct 18, 2009 at 5:07 PM, Adam Kocoloski <[email protected]> wrote:
On Oct 17, 2009, at 9:15 PM, Jan Lehnardt wrote:

Hey all,

I had an itch to scratch and now there is an experimental branch that
enables logging to a database:

 http://github.com/janl/couchdb/commits/log-to-db

Check out the test case on how it is supposed to work:


http://github.com/janl/couchdb/blob/5cbfbeaddea2152e35b88cae515152f78d6b7544/share/www/script/test/log_database.js

You just create a `database = dbname` entry in your configuration under the `[log]` section and CouchDB will start to write log entries to the
specified database.

Depending on the log level this means that each single request to CouchDB triggers multiple writes to the log database. The logger, for now, defaults to delayed_commits. A future improvement, could be using the `batch=ok`
module.

I just wanted to throw this out here inspired by Chris' "releasing things
before I go to sleep for the night".

Let me know what you think.

Cheers
Jan
--

Hi Jan, cool!  I do have a word of caution, though ...

I tried this a long time ago but never released it because it caused the logging queue to back up pretty badly under a moderate read load. Logging events are asynchronous (as they should be), so if the logging event handler is slower than the system generating the events its entirely possible for the message queue in the gen_event manager to back up. Unfortunately, Erlang has this ugly positive feedback loop where a large message queue will dramatically slow down that process. Eventually (assuming a constant read load), the gen_event manager process used all available memory and things
went kaboom.

maybe using gen_event2 from rabbitmq project would solve that since it buffer ?
http://hg.rabbitmq.com/rabbitmq-server/file/4bb959b80bac/src/gen_server2.erl

- benoit

Is there a gen_event2? That link goes to gen_server2. In principle a gen_event2 would help, if you knew that the load on the server would eventually drop to the point that the logger could catch up. I'd be fairly nervous about betting on that in production, though. Best,

Adam

Reply via email to