Show HN: Silk, a profiling tool for Django(mtford.co.uk)
mtford.co.uk
Show HN: Silk, a profiling tool for Django
http://mtford.co.uk/blog/2/
24 comments
Probably not a good idea to keep your profiler open to the world as long as you allow reading from arbitrary files. After a bit of fiddling around I was able to read your secret key from the settings file. It starts with 'es!b3'. And your soundcloud password. Sorry.
This is a really great tool, massive fan, made some good use of it in the ~15 hours since I installed it yesterday evening. Are you aware it seems to bork the admin interface?
If, after installing Silk, I use admin to update a record, I get "You cannot access body after reading from request's data stream" - this applies across all admin models and pages. It's not really that big a deal, as I rarely use the admin pages now, but it might have knock-on effects that I haven't seen yet. (Commenting out Silk in the installed apps list clears the problem up straight away.)
I'll ping you an email with the stack trace.
If, after installing Silk, I use admin to update a record, I get "You cannot access body after reading from request's data stream" - this applies across all admin models and pages. It's not really that big a deal, as I rarely use the admin pages now, but it might have knock-on effects that I haven't seen yet. (Commenting out Silk in the installed apps list clears the problem up straight away.)
I'll ping you an email with the stack trace.
I wrote https://github.com/rory/django-sql-inspector a django app to profile SQL queries done in multiple pages, including showing stack trace of where it's being called. You can find out which function(s) are responsible for long SQL requests.
What is the overhead (roughly) for this guy? I do absolutely love the idea of using elasticsearch for this instead of a traditional database.
Using celery or django-rq to delay the processing also seems reasonable.
Using celery or django-rq to delay the processing also seems reasonable.
Tbh, probably fairly bad atm. I haven't done anything empirical but atm Silk will create a record for every SQL query executed during the request/response cycle effectively doubling the number of queries. Not only that but it also saves down any non-binary HTTP body to a TextField, and this isn't yet configurable.
I think the Elasticsearch/Celery option is a good shout if this were to ever be used in production. Celery would help with the issue of response time but it wouldn't solve the load impact on the database (although I guess could also configure a different SQL database in Django).
I will certainly do some investigation into this at some point :)
I think the Elasticsearch/Celery option is a good shout if this were to ever be used in production. Celery would help with the issue of response time but it wouldn't solve the load impact on the database (although I guess could also configure a different SQL database in Django).
I will certainly do some investigation into this at some point :)
StatsD might be better since you are just recording datapoints + text field I'm assuming.
https://pypi.python.org/pypi/python-statsd
https://pypi.python.org/pypi/python-statsd
Sounds interesting... i'll take a look. The fact that its UDP would be helpful in avoiding the need for a dependency on celery/message queues.
What's it like in terms of getting the data back out?
What's it like in terms of getting the data back out?
You can write your plugin for the back end to dump it. It defaults to graphite which would be familiar enough I think.
For people interested in production monitoring I've created https://appenlight.com - and our python middleware will integrate and start timing django applications without any code changes required (except 3 lines to load middleware itself).
@mtford great job!
@mtford great job!
Very cool, this seems much nicer than using hotshot to profile Django.
I'm a little confused by this page though: http://mtford.co.uk/silk/request/1907/profiling/. Why is wrapped_target listed so many times for a single blog post query? Is there a hierarchy to the profile calls that's not shown? Oh, does it go deepest level to highest level calls?
I'm a little confused by this page though: http://mtford.co.uk/silk/request/1907/profiling/. Why is wrapped_target listed so many times for a single blog post query? Is there a hierarchy to the profile calls that's not shown? Oh, does it go deepest level to highest level calls?
Thats a great question... that def shouldn't be there. render_to_response is actually dynamically profiled i.e. the decorator is applied at runtime via configuration in settings.py. This looks like its probably a bug with that. Cheers, will look into it
After reading through this intro, it isn't clear to me if I can use the decorator or context manager to profile code outside the request/response cycle.
Any ideas if that's the case?
Any ideas if that's the case?
Not at the moment. Silk doesn't actually save anything down until right at the end of the cycle i.e. in process_response of the middleware. Had to do this to avoid issues with atomic transactions. I certainly don't mind adding this as a feature if people would find that useful - perhaps by detecting that no request is in process and then commiting profile data on the fly if that's the case.
absolutely great tool!
Looks very interesting! We currently use django debug toolbar for profiling, which seems to have a lot of overlap. I like how Silk stores the profiles of multiple requests - that seems to be a nice differentiator. What else? Thanks for the awesome contribution to the Django community!
Yep django-toolbar is great! My aim with this was, like you said, to provide the history of multiple requests, but also to provide finer-grained profiling with the decorator and context manager. Not only do they capture execution time but they also capture the queries that were executed whilst they were active. The aim was extra visibility all round really.
> I like how Silk stores the profiles of multiple requests
You might like my Django SQL Inspector which aggregates many SQL queries across many requests.
https://github.com/rory/django-sql-inspector
You might like my Django SQL Inspector which aggregates many SQL queries across many requests.
https://github.com/rory/django-sql-inspector
Looks really nice.
Have you thought that instead of sending the results to elasticsearch? Instead of using something like newrelic this could take it's place..
Have you thought that instead of sending the results to elasticsearch? Instead of using something like newrelic this could take it's place..
That's not a bad idea, it would probably be more natural than the rigid filters that are currently in place for the requests and profiles. As far as having this running in prod I've only really thought of siphoning off the results to celery or something similar to avoid impacting response time, but a spray and pray at Elasticsearch would work just as well.
What's nice about the current setup is that it takes very little effort to get up and running but I suppose could make that configurable...
What's nice about the current setup is that it takes very little effort to get up and running but I suppose could make that configurable...
If you are looking for something like newrelic you might want to give https://appenlight.com a try.
I've put a lot of love into it and it can give you tons of information about your application.
I've put a lot of love into it and it can give you tons of information about your application.
Thanks for posting this! Could you add a download link somewhere?
Sure just added this to the intro. You can download via https://github.com/mtford90/silk/releases or install through pip install django-silk