Opened 9 months ago
Last modified 6 months ago
#35167 assigned Bug
JSONField get_db_prep_value being called with `Cast` types
Reported by: | Samantha Hughes | Owned by: | HeejunShin |
---|---|---|---|
Component: | Database layer (models, ORM) | Version: | 4.2 |
Severity: | Normal | Keywords: | JSONField |
Cc: | Samantha Hughes, Simon Charette, Krish, Jacob Walls | Triage Stage: | Accepted |
Has patch: | no | Needs documentation: | no |
Needs tests: | no | Patch needs improvement: | no |
Easy pickings: | no | UI/UX: | no |
Description
Running django 4.2.9.
We have a customized version of the standard JsonField
where we have overridden from_db_value
and get_prep_value
.
The worked in django ~4.1. Migrating to 4.2.9, the get_prep_value
function started receiving Cast
objects as the value
, exploding json serializing/encryption , only during bulk_update
.
We fixed it by overriding `get_db_prep_save to
def get_db_prep_save(self, value, connection): if hasattr(value, "as_sql"): return value return self.get_db_prep_value(value, connection=connection, prepared=False)
The jsonfield implementation does not check for as_sql
. I'm not sure if this is intentional or not (unmodified JSONField columns don't seem to explode), possibly our implementation has messed up the jsonfield?
Change History (24)
comment:1 by , 9 months ago
Triage Stage: | Unreviewed → Accepted |
---|---|
Type: | Uncategorized → Bug |
comment:2 by , 9 months ago
You mentioned that you are using Django 4.2 so the following patch won't apply cleanly against your branch but against main it gets the tests passing
-
django/db/models/fields/json.py
diff --git a/django/db/models/fields/json.py b/django/db/models/fields/json.py index 571e6e79f3..73b4a4dd50 100644
a b def get_internal_type(self): 99 99 def get_db_prep_value(self, value, connection, prepared=False): 100 100 if not prepared: 101 101 value = self.get_prep_value(value) 102 if isinstance(value, expressions.Value) and isinstance(103 value.output_field, JSONField104 ):105 value = value.value106 elif hasattr(value, "as_sql"):107 return value108 102 return connection.ops.adapt_json_value(value, self.encoder) 109 103 110 104 def get_db_prep_save(self, value, connection): 105 # This slightly convoluted logic is to allow for `None` to be used to 106 # store SQL `NULL` while `Value(None, JSONField())` can be used to 107 # store JSON `null` while preventing compilable `as_sql` values from 108 # making their way to `get_db_prep_value` which is what the `super()` 109 # implementation does. 111 110 if value is None: 112 111 return value 113 return self.get_db_prep_value(value, connection) 112 elif ( 113 isinstance(value, expressions.Value) 114 and value.value is None 115 and isinstance(value.output_field, JSONField) 116 ): 117 value = None 118 return super().get_db_prep_save(value, connection) 114 119 115 120 def get_transform(self, name): 116 121 transform = super().get_transform(name)
Basically now that the deprecation shims have been removed the logic can be slightly simplified to only have to special case the Value(None, JSONField())
which is truly the only problematic case in the implementation of JSONField
due to ambiguous nature of None
with regards to SQL NULL
and JSON null
.
comment:3 by , 9 months ago
Cc: | added |
---|
comment:4 by , 9 months ago
Thanks, the quick response is really appreciated. Any idea if this will make it to a 4.2 patch or will we need to upgrade to ~5?
comment:5 by , 9 months ago
Component: | Error reporting → Database layer (models, ORM) |
---|---|
Owner: | set to |
comment:6 by , 9 months ago
Any idea if this will make it to a 4.2 patch or will we need to upgrade to ~5?
Assuming someone prepares a patch it should make the cut for 5.1 meant to be released in August 2024. It won't be backported to 4.2 or 5.0 at this point unfortunately.
comment:7 by , 9 months ago
Cc: | added |
---|
comment:8 by , 9 months ago
Cc: | added |
---|---|
Summary: | JSONFIeld get_db_prep_value being called with `Cast` types → JSONField get_db_prep_value being called with `Cast` types |
comment:10 by , 9 months ago
@HeejunShin, nobody owns this ticket, so you're welcome to work on this ticket. You can assign it to yourself and/or add comments as you work on it.
comment:11 by , 9 months ago
Owner: | changed from | to
---|---|
Status: | new → assigned |
comment:12 by , 8 months ago
Is this a relevant problem?
Because, after some period of exploration, I see that most of the code were updated.
comment:13 by , 8 months ago
Owner: | removed |
---|---|
Status: | assigned → new |
comment:15 by , 8 months ago
Replying to HeejunShin:
Is this ticket solved?
No. While it has an inline tentative patch no one has opened a PR on Github with the proposed changes including a regression test yet.
comment:18 by , 8 months ago
Owner: | set to |
---|---|
Status: | new → assigned |
comment:19 by , 8 months ago
Replying to Simon Charette:
thank you for youre response.
comment:20 by , 7 months ago
Sorry for the slow progress. I am currently analyzing the code looking at the customizing doc of the Field class.
comment:22 by , 7 months ago
Replying to Natalia Bidart:
Ticket #35381 is related to this.
Is this a problem that needs to be resolved as quickly as possible?
Should I also refer to the ticket you mentioned to solve the problem?
Any advice would be greatly appreciated.
follow-up: 24 comment:23 by , 7 months ago
Hello HeejunShin, thank you for asking! You can take the time you need to work on this ticket, in general no ticket is urgent unless is a release blocker, which is not the case for this one.
Also, if you can't work on this anymore, that is also fine and you can just de-assign yourself from the ticket and eventually someone else would pick it up.
Django is built and maintained by its community and everyone contributes their time and energy when they can.
comment:24 by , 7 months ago
Replying to Natalia Bidart:
Hello HeejunShin, thank you for asking! You can take the time you need to work on this ticket, in general no ticket is urgent unless is a release blocker, which is not the case for this one.
Also, if you can't work on this anymore, that is also fine and you can just de-assign yourself from the ticket and eventually someone else would pick it up.
Django is built and maintained by its community and everyone contributes their time and energy when they can.
thank you for your response.
I'll try my best.
comment:25 by , 6 months ago
It's really late though
I was just analyzing the code and reproduced the bug.
comment:26 by , 6 months ago
i have a question.
maybe The bug reproduction seems to be incorrect.
i tested using 2 ways.
tests[1].json = {"test": Cast('good', output_field=CharField())} Test.objects.bulk_update(tests_will_be_updated, fields=["json"])
tests[1].json = Cast("good", output_field=CharField()) Test.objects.bulk_update(tests_will_be_updated, fields=["json"])
first one, i got "raise TypeError(f'Object of type {o.class.name} 'TypeError: Object of type Cast is not JSON serializable" error.
but value argument type in get_db_prep_value is dict.
second one, i could not get error.
value argument type in get_db_prep_value is <class 'django.db.models.expressions.Case'>. (CASE WHEN <WhereNode: (AND: Exact(Col(polls_test, polls.Test.id), 2))> THEN Cast(Col(polls_test, polls.Test.json)), ELSE Value(None)).
i thought first one is type dict, so this is wrong.
is second one wrong way that reproduce bug case?
This highly likely relates to #34539 which was about making sure
get_db_prep_value
always callsget_prep_value
as it use to be case in 4.1.Given the default implementation of
get_db_prep_save
doesn't delegate toget_db_prep_value
when provided a compilable expression I believe we should also make sureJSONField.get_db_prep_save
does something equivalent.The challenge here is that
Value(output_field=JSONField())
, which represents wrapped literal JSON values which could also be JSON strings, must be able to make their way toconnection.ops.adapt_json_value(value, self.encoder)
. At least they needed during the deprecation period maybe it's no longer the case sinceValue(value).as_sql
will do the right thing by itself.