Opened 16 months ago
Last modified 3 months ago
#34746 assigned Cleanup/optimization
High CPU/memory consumption when a 5XX is raised with large local variables
Reported by: | Rémi Dupré | Owned by: | Keerthi Vasan S A |
---|---|---|---|
Component: | Error reporting | Version: | 4.2 |
Severity: | Normal | Keywords: | |
Cc: | Triage Stage: | Accepted | |
Has patch: | yes | Needs documentation: | no |
Needs tests: | yes | Patch needs improvement: | yes |
Easy pickings: | yes | UI/UX: | no |
Description (last modified by )
Hi!
In a view with large variable in its scope, if an exception is raised, the worker will freeze and its memory usage will raise. Here is a minimal example:
import sys from django.urls import path # 300MB large_data = list(range(40 * 1024 * 1024)) def compute_lines(request): local_data = large_data raise ValueError urlpatterns = [ path("compute-lines/", compute_lines) ]
When calling /compute-lines/ you will notice the issue : the server will take a few seconds to responds while its memory raises by ~1GB.
After a bit of investigations, it turned out it comes from django.utils.log.AdminEmailHandler in the default logging config. This handler will eventually pretty print the whole context of each trace frame from here : https://github.com/django/django/blob/main/django/views/debug.py#L355
This implementation ensures that no more than 4kB are included in the formatted result, but the entire variable will still have to be rendered first. I'm not sure how it could be solved, but maybe https://docs.python.org/3/library/reprlib.html can be a clue?
Change History (20)
comment:1 by , 16 months ago
Description: | modified (diff) |
---|---|
Summary: | High CPU/memory consumtion when a 5xx is raised with a large local variable → High CPU/memory consumption when a 5XX is raised with large local variables |
comment:2 by , 16 months ago
Easy pickings: | set |
---|---|
Triage Stage: | Unreviewed → Accepted |
Type: | Bug → Cleanup/optimization |
comment:3 by , 16 months ago
After some experimentation following the suggested link to reprlib, perhaps something like this could be a good alternative? (it needs better code formatting, more reprlib limits sets and tests, of course)
-
django/views/debug.py
a b class ExceptionReporter: 347 347 self.template_does_not_exist = True 348 348 self.postmortem = self.exc_value.chain or [self.exc_value] 349 349 350 import reprlib 351 repr_instance = reprlib.Repr() 352 repr_instance.maxstring = 5000 353 repr_instance.maxlist = 1000 354 350 355 frames = self.get_traceback_frames() 351 356 for i, frame in enumerate(frames): 352 357 if "vars" in frame: 353 358 frame_vars = [] 354 359 for k, v in frame["vars"]: 355 v = pprint(v)360 v = repr_instance.repr(v) 356 361 # Trim large blobs of data 357 362 if len(v) > 4096: 358 363 v = "%s… <trimmed %d bytes string>" % (v[0:4096], len(v))
comment:4 by , 16 months ago
Owner: | set to |
---|---|
Status: | new → assigned |
comment:5 by , 16 months ago
Owner: | removed |
---|
comment:6 by , 16 months ago
Owner: | set to |
---|
comment:8 by , 16 months ago
I think using reprlib is acceptable, this is exactly the kind of situation it’s designed for. It will be a shame to lose the pretty-printing aspect, but perhaps reprlib/pprint will grow appropriate functionality in the future.
comment:9 by , 12 months ago
Owner: | changed from | to
---|
Hello folks, I saw the two pull requests that were attached in the ticket are stale, and the original author have both closed. I'm interesting in picking this as my first ticket, can I proceed?
comment:10 by , 12 months ago
I was:
- able to verify the following
- when exception is raised after declaring a large variable, memory usage shoots up.
- when exception is not raised after declaring a large variable, memory remains in live with normal operations.
Setup I used to reproduce the error and suggested solution
- added a middleware which logs the change in memory after each request.
- created two views, one which throws error and one that does not.
- observed the memory usage in both the cases.
Current
After using reprlib
comment:11 by , 10 months ago
Owner: | changed from | to
---|
comment:12 by , 10 months ago
Triage Stage: | Accepted → Ready for checkin |
---|
comment:13 by , 10 months ago
Triage Stage: | Ready for checkin → Accepted |
---|
You should not mark your own PRs as "ready for checkin".
follow-up: 15 comment:14 by , 10 months ago
Ah sorry, I wanted to set this ticket as ready to be reviewed as mentioned by a contributor in the PR. How do I do that?
comment:15 by , 9 months ago
Replying to Keerthi Vasan S A:
Ah sorry, I wanted to set this ticket as ready to be reviewed as mentioned by a contributor in the PR. How do I do that?
Hi!
The proper procedure to get a re-review is described in the PR review checklist. Basically you need to unset the relevant flags in this ticket so the PR gets added back to the "patches needing review" section of the Django Developer Dahsboard.
comment:17 by , 9 months ago
Needs tests: | unset |
---|---|
Patch needs improvement: | unset |
comment:18 by , 9 months ago
Needs tests: | set |
---|---|
Patch needs improvement: | set |
comment:20 by , 3 months ago
Replying to Ahmed Ibrahim:
Is this still being worked on or should I handle t?
Hello Ahmed! Thanks for your eagerness to help. As we already discussed in the PR, this is still being work on and I'll provide more feedback in the PR.
I have managed to reproduce the issue, thank you for the example. Indeed if
v = pprint(v)
is replaced by something that does not render the whole variable, the memory and CPU usage do not spike.Accepting so we can evaluate whether an alternative could be used instead of
pprint
that would account for the size limit before printing.