Startled by the observation that Hyperkitty's unit tests to fail because an excessive amount of file descriptors are opened, I began to dig a little.
The majority of these file descriptors are opened by "Django Q" -- a django library for asynchronous tasks.
I wondered what these asynchronous tasks might be in an e-mail archiver. After all, it has to perform 2 very specific jobs:
- Receive e-mails via http requests, process them, and store them in a database.
- Display the collected e-mails.
On top of that, it generates some statistics, includes a voting system, and allows replying to e-mails in a forum-like fashion.
Job 1 really requires very little processing and is mostly database IO as the message has to be sorted and its position in the thread has to be computed. None of this can be efficiently performed in the background/asynchronously because there is nothing to be parallelized. Performing the *whole* receive-and-store operation non-blocking also is not an option -- ultimately mailman needs to know if it succeeded.
Job 2 is more or less what every web application does (reading from a database and rendering the result) and certainly does not require any asynchronous processing.
So where is Django Q used in Hyperkitty? It all boils down to the file hyperkitty/tasks.py where a smaller number of asynchronous tasks are defined. They can be grouped into 3 classes:
- query mailman core
- repair the data structure (empty threads, orphans, ...)
- rebuild query cache
I would like to argue that none of these 3 groups of tasks need to be performed asynchronously.
- Mailman-core only needs to be queried when mailing lists are changed. This is triggered by signals from postorius and in addition periodically by a cron job.
- The data structure should not need to be repaired in the first place but the appropriate on_delete/on_save triggers should take care of this -- and I believe they do in recent versions. If for some reason the database becomes corrupted one can always start a repair operation. Nothing is gained however, from running this asynchronously.
- Lastly the cache rebuild: Currently Hyperkitty rebuilds its cache (which caches the db queries, not the frontend) whenever an e-mail is received. Since it only involves *reading* from the db it actually is something that *can* be done asynchronously to reduce the time it takes to process an incoming e-mail.
But is it really worth the tremendous additional complexity that is introduced by Django Q?
- requires a "qcluster" to run in the background (see shipped unit file)
- loss of determinism / debugging becomes much harder
- enourmous amount of file descriptors are opened in testing
- additional dependency
A similar result can be obtained by simply scaling the wsgi application accordingly (if needed) and/or optimizing the db queries. Alternatively one could simply invalidate the affected caches instead of rebuild them every time an e-mail is received or don't trigger cache rebuilds on received e-mails at all...
But maybe I overlooked something. I argue that we do not *need* Django Q. The question is: do we want it? What are your thoughts on this?
Best Leon