Opened 16 years ago
Closed 12 years ago
#10200 closed Bug (fixed)
loaddata command does not raise CommandError on errors
Reported by: | Lorenzo Gil Sanchez | Owned by: | nobody |
---|---|---|---|
Component: | Core (Management commands) | Version: | dev |
Severity: | Normal | Keywords: | exit status |
Cc: | yishaibeeri | Triage Stage: | Accepted |
Has patch: | yes | Needs documentation: | no |
Needs tests: | no | Patch needs improvement: | no |
Easy pickings: | no | UI/UX: | no |
Description
When there is an error in a fixture being loaded by the command loaddata it prints an error description into stderr and them simple return. I think it would be better to raise a CommandError instead of simple exiting the handle method. My reasons are:
- You can't easily know the succesfullnes of a loaddata command programmatically. For example, we have a buildbot that load some fixtures and we don't know if it passes this step or not since the exit code is always 0 no matter the data is successfully loaded or not.
- AFAIK is the standard Django mechanism to report command errors
- The code that catch this type of exceptions does already print its message to stderr, so we do not lose that functionality.
I'm attaching a patch that fixes this issues with existing tests already fixed.
Attachments (2)
Change History (18)
by , 16 years ago
Attachment: | scream-if-error-in-loaddata-command.diff added |
---|
comment:1 by , 16 years ago
Has patch: | set |
---|
comment:2 by , 16 years ago
milestone: | post-1.0 |
---|
comment:3 by , 16 years ago
Triage Stage: | Unreviewed → Design decision needed |
---|
This is a backwards-incompatible change, so we need a very good reason to make this change. I'm not sure your reasons get that far, but I'll leave this open for some more discussion.
comment:4 by , 16 years ago
What harm could do this patch to users other than alerting them that they used to have bad fixtures and now they know it?
The only real difference this patch introduces is that loaddata stops trying to load fixtures as soon as one of them is rotten. I think this is a good thing since the previous behaviour could hide obscure bugs in your data.
comment:5 by , 16 years ago
I noticed what is likely a related problem. When the loaddata
command is used with the auto-generated manage.py, bad syntax
results in silent failure and a return code indicating success.
The built-in usage help shows that the fixture argument is
mandatory:
Usage: manage.py loaddata [options] fixture [fixture ...]
However, manage.py does not report an error when a fixture is not
specified:
$ ./manage.py loaddata $ echo $? 0
Only when the verbosity level is increased from 0 to 2 is an error
reported, and the return code still indicates success:
$ ./manage.py loaddata --verbosity=1 ; echo $? 0 $ ./manage.py loaddata --verbosity=2 ; echo $? No fixtures found. 0
Typically, silence implies successful completion. This problem is
particularly troublesome because the syntax for loaddata differs
significantly from that of dumpdata. It took
outside assistance
for me -- new to Django but with a considerable amount of experience
using CLI utilities -- to determine why the following commands
completed with apparent success but left my database empty:
$ ./manage.py dumpdata >data.json $ ./manage.ph flush $ ./manage.py loaddata <data.json
comment:6 by , 16 years ago
Cc: | added |
---|
comment:7 by , 16 years ago
Cc: | removed |
---|
(noticed I receive e-mail updates after commenting regardless of having my address in the CC field)
comment:8 by , 14 years ago
Severity: | → Normal |
---|---|
Type: | → Bug |
comment:9 by , 14 years ago
Patch needs improvement: | set |
---|
comment:10 by , 13 years ago
Easy pickings: | unset |
---|---|
Resolution: | → duplicate |
Status: | new → closed |
UI/UX: | unset |
#16026 would solve this one in a quite elegant way. Marking this one a duplicate of that.
comment:11 by , 13 years ago
Cc: | added |
---|---|
Component: | Core (Serialization) → Core (Management commands) |
Keywords: | exit status added |
Resolution: | duplicate |
Status: | closed → reopened |
Triage Stage: | Design decision needed → Accepted |
Version: | 1.0 → SVN |
Reopening, as the fix for #16026 did not, eventually, fix this issue. In trunk (as of rev 17029) the exit status of a failed loaddata invocation (e.g., bad value in a fixture causing a DatabaseError) is still 0.
comment:12 by , 13 years ago
Needs tests: | set |
---|
Actually, loaddata code in its current status is such that all error code paths lead to:
- An error is printed to stderr
- The handle() method returns
So, the only (beneficial IMHO) change the modification proposed by this ticket would introduce would be that the Unix process exit code changes from zero to a non-zero value.
I suspect the backward-incompatibility reservations Jacob had back then were related to the fact that raising CommandError would mean exiting from the nested loops and aborting execution of the command; but the evolution of the code during the last three years has resulted in precisely that behavior already
But I'm not sure. Will try to confirm with him.
comment:13 by , 13 years ago
Any news about this ?
I've just hit this. We have a buildbot and when the loaddata fails, we cannot easily detect it. So, having run "set -e" in my bash doesn't help in this case, and then tests are run. The tests fail and the error was that loaddata failed, very hidden in the output.
If the exit code of the command could be just like any standard unix command (!=0 on error), that will be great and made life for sysadmins like me happier :)
Btw, we are using django 1.3.1 in case it matters
comment:14 by , 12 years ago
When #18387 will be committed, we should be able to easily make a clean patch for this one.
comment:15 by , 12 years ago
Needs tests: | unset |
---|---|
Patch needs improvement: | unset |
Here's an updated patch for this issue.
comment:16 by , 12 years ago
Resolution: | → fixed |
---|---|
Status: | reopened → closed |
Milestone post-1.0 deleted