Databricks SQL script slow execution in workflows using serverless

[iamgoda]

I am running a very simple SQL script within a notebook, using an X-Small SQL Serverless warehouse (that is already running). The execution time is different depending on how it's run:

• 4s if run interactively (and through SQL editor)

• 26s if run within a workflow as a notebook task

If you look at the query history screenshot below you see the individual queries are taking the same amount of time to execute, but for some reason there's a few seconds delay between end of one query and start of the next query (line) when running through workflow vs interactively (both using SQL serverless warehouse). I tried both current and preview channel and they both behave the same way

I even tried using an all purpose compute cluster (Single node: Standard_DS3_v2 · DBR: 15.3) and the times were consistent when running in a workflow vs interactively.

Script

declare or replace start_time timestamp = current_timestamp();
declare or replace end_time timestamp = current_timestamp();
declare or replace v STRUCT<SOURCE_TBL_MAX_TIMESTAMP STRING, TARGET_TBL_MAX_TIMESTAMP STRING, SOURCE_MAX_TIMESTAMP TIMESTAMP, TARGET_MAX_TIMESTAMP TIMESTAMP, MIN_TIME TIMESTAMP, MAX_TIME TIMESTAMP, MAX_INTERVAL_TIME TIMESTAMP, TABLE_SOURCE STRING, TABLE_TARGET STRING, MERGE_STATEMENT STRING, INTERVAL_COLUMN_NAME STRING, STAGING_TABLE_NAME STRING, STAGING_TABLE_CREATE_STATEMENT STRING, FULL_STAGING_TABLE_NAME STRING>;
select current_timestamp();
select v;
select current_timestamp();
select timestampdiff(MILLISECOND  , start_time, current_timestamp()) as total_time_elapsed_ms;

Screenshots 

Query History

Further tests

Using a SQL file workflow task, running on X-Small SQL serverless, the same script executes as expected (faster than notebook task):

Not sure why the type of workflow task would impact the execution of the SQL script on a serverless warehouse.

[iamgoda]

@SergeRielau - is this another bug or?

[SergeRielau]

@iamgoda No idea. In DBSQL session variables and their values must be preserved outside the runtime engine (since the session may be migrated), but that doesn't seem to be at play here (and it should not have such an impact).

[iamgoda]

Yes and the fact that running it as a SQL script task works. Running it as a notebook task on a SQL serverless instance seems to be the issue.

[BilalAslamDbrx]

@iamgoda  we are going to look into how to make this faster. There's a poll loop in Databricks Workflows for SQL notebooks (but not for SQL scripts) which causes things to slow down. 

[iamgoda]

@BilalAslamDbrx thanks for looking into this and confirming the issue. This doesn't happen if the SQL notebook task is run using all purpose compute in a workflow, so wondering if the poll loop only applies to SQL serverless compute? 

[BilalAslamDbrx]

This adds up. This particular polling inefficiency only applies to the flow of notebooks running on warehouses (not on clusters).

[iamgoda]

Any idea of whether this will be addressed and when?

[BilalAslamDbrx]

No not yet

[Ememel]

Hello. Are there updates for this issue?

[iamgoce]

I was told that it would be fixed in Q4 this year. I am checking for an update again and will let you know if they give me a more exact date.

[iamgoce]

So I was told that the Q4 date was incorrect - in fact there is currently no ETA for when this issue will be fixed. It's considered lower priority by Databricks as not enough customers are impacted or have raised this type of an issue. I would recommend @Ememel that you try and raise this issue through your Databricks support team or Microsoft so it increases the priority?