Repository navigation
SQLColumnCheckOperator failures after upgrading to common-sql==1.3.0 #27976
Description
Activity
- addedkind:bugThis is a clearly a bugThis is a clearly a bug
on Nov 28, 2022 Thanks for opening your first issue here! Be sure to follow the issue template!
@potiuk seems like there is another dependency issue here. I noticed the
DBApiHookchanged some behavior whereget_all()is now callingrun()to make these calls. Not sure if that's the issue or maybe something Snowflake-specific, but the return type of the record has changed.@mag3141592 the
.get("tolerance")is fine as it is, as thevalueparam defaults toNone. All previous versions of the operator have been written like this and they work fine!Seems also we have the same (or similar) problem in #27978
We might need a 1.3.2 rather quickly I am afraid.
Interesting. Is this at the hook level and something that needs to be changed for the operators? If so, I can open a PR today with the fix for a dict return type.
Still not sure if this is a Snowflake-specific problem or a general SQL hook issue. It doesn't seem like
run()is necessarily returning a dict.I do not know yet :) . I am travelling most of today and I had no time to look closely as I was releasing 1.3.1 + other providers.. I will see later today but maybe you can see in the meantime what would be the possible reason.
Reacted by Benji Lampel- addedpriority:mediumBug that should be fixed before next release but would not block a releaseBug that should be fixed before next release but would not block a release
on Nov 29, 2022 We had the same issue on the Astro SDK side, which we had to fix by astronomer/astro-sdk@0c64f48
Reacted by Benji LampelWe had the same issue on the Astro SDK side, which we had to fix by astronomer/astro-sdk@0c64f48
Thanks @kaxil - I think the root cause of this problem is different. This was the problem of the "unifying" the run() methods
behaviour I did in #27912.Fix for the problem you found in astro is here: #27997
I am still looking for the other problems - they look similar but they were introduced in 1.3.0 so this is not the case.
OK. @denimalpaca @kazanzhy.
I know why it is happening - it is Snowflake-only thing. And I am not 100% sure yet how to fix it in a good way.
The root cause of the problem is this line in Snowflake Hook's run() method:
with closing(conn.cursor(DictCursor)) as cur:The result of it is that after the #26944 change, get_records() and get_first() changed the result type to return dictionaries and not sequences (previously each of those methods had their own implementations and did not use run() method, so they used "standard" sensor).
Simply speaking: SnowflakeHook's run() method is not standard - instead of sequence of sequences, it returns sequence of dicts. This is absoluately not standard behaviour for DBApI - this is simply creative aproach of snowflake python connector to return Dicts: https://docs.snowflake.com/en/user-guide/python-connector-api.html#cursor
I think the best fix is:
- as of common-sql 1.3.1 we don not need to use DictCursor any more. We keep ".description" field in the hook and we can
use it to convert the return value in Operator from Sequences to Dics (see below) - we standardise SnowflakeHook's run() method to return sequences by default (breaking change - but only for "deep" usages for where hook is used by custom operators or taskflow cases)
- we keep the SnowflakeOperator behaviour to return Dicts (when do_xcom_push = True) to avoid breaking changes at "task dependency" level (we can easily convert the output by combining description + results from the Hook into return value of the SnowflakeOperator.
This would be almost identical change to what I've done for Databricks last week (only Databricks internals were different - no DictCursor and it returned Tuples rather than Dicts):
The DatabricksSqlHook is now conforming to the same semantics as all the other DBApiHook implementations and returns the same kind of response in its run method. Previously (pre 4.* versions of the provider, the Hook returned Tuple of (“cursor description”, “results”) which was not compatible with other DBApiHooks that return just “results”. After this change (and dependency on common.sql >= 1.3.1), The DatabricksSqlHook returns now “results” only. The description can be retrieved via last_description field of the hook after run method completes.
That would require Snowflake 5.0.0 breaking change release, and basically making whole 4.0.* yanked. Or probably better - we could relase 4.0.2 with adding this breaking change and yanking 4.0.0, 4.0.1 as "wrong". Making 4.0.2 the "real breaking" change we actually wanted (or started to want) to make in 4.0* . 4.0.* is very fresh and has not been released yet in constraints.
The end result will be that we will finally (I hope) standardise the semantics of DBApiHook.run(). So far, each of the hooks could change it (and some did as we see) - this change seems to finally got all the run() methods to behave in the same way.
WDYT Everyone?
Reacted by Dmytro Kazanzhy- as of common-sql 1.3.1 we don not need to use DictCursor any more. We keep ".description" field in the hook and we can
The same fix will work for #27978 - standardising run() methods to behave the same for all operators will also fix #27978, because SqlSensor uses
get_records()and gets sequence of dicts instead of sequence of sequences and 0 is missing in the dictionary.CC: @uranusjr @kaxil I am also summoning you for opinion on that one
This is how the fix can look like #28006 - so it is not big. If others will think this is a good, idea (@mik-laj - maybe you also can take a look and see if that's ok to add this breaking change) then I will add similar tests as we have in Databricks Hook now https://github.com/apache/airflow/blob/main/tests/providers/databricks/hooks/test_databricks_sql.py
Those involved - Let me know please if you think this is a good idea to add this breaking change (cc: @eladkal ?)
I've added full tests coverage in #28006
My vote is towards standardizing Snowflake's hook with other DBApi Hooks, and having 4.0.2 being "the" breaking change for deep usages.
Reacted by Jarek PotiukMy vote is towards standardizing Snowflake's hook with other DBApi Hooks, and having 4.0.2 being "the" breaking change for deep usages.
Yeah I think it makes most sense.. This is one-time pain we have to go throug to finally get "common" behaviours and be able to use it in lineage, common.sql operators, transfers and a number of possible future usages.
And users can still use 3.* for a long time if they want to keep compatibility
- added a commit that references this issue
on Dec 1, 2022 - added a commit that references this issue
on Dec 1, 2022 @mag3141592 - can you please check if 4.0.2rc1 solves the problem and report it in #28036
Apache Airflow Provider(s)
common-sql
Versions of Apache Airflow Providers
apache-airflow-providers-google==8.2.0
apache-airflow-providers-http==4.0.0
apache-airflow-providers-salesforce==5.0.0
apache-airflow-providers-slack==5.1.0
apache-airflow-providers-snowflake==3.2.0
Issue:
apache-airflow-providers-common-sql==1.3.0
Apache Airflow version
2.4.3
Operating System
Debian GNU/Linux 11 (bullseye)
Deployment
Astronomer
Deployment details
No response
What happened
Problem occurred when upgrading from common-sql=1.2.0 to common-sql=1.3.0
Getting a
KEY_ERRORwhen running a unique_check and null_check on a column.1.3.0 log:

1.2.0 log:

What you think should happen instead
Potential causes:
COL_NAMEinstead of the table columnSTRIPE_IDrecordfrom the test changed types went from a tuple to a list of dictionaries.toleranceis specified for these tests, so.get('tolerance')looks like it will cause an error without a default specified like.get('tolerance', None)Expected behavior:
toleranceis not a required key.How to reproduce
Anything else
No response
Are you willing to submit PR?
Code of Conduct