Automatically configure BigQuery scan parallelism #22279
Conversation
findepi
force-pushed
the
findepi/bq-auto
branch
3 times, most recently
from
06b922a to
0437630
Compare
plugin/trino-bigquery/src/main/java/io/trino/plugin/bigquery/BigQueryConfig.java
Outdated
Show resolved
Hide resolved
plugin/trino-bigquery/src/main/java/io/trino/plugin/bigquery/BigQuerySplitManager.java
Outdated
Show resolved
Hide resolved
There was a problem hiding this comment.
@findepi Did you mean to clamp this to 100 at-most? Current code is min(worker * 3, 100) which can be less than 100.
Did you mean max(100, min(worker * 3, 100))?
Or change the comment to Limit to 100 for very large clusters.
EDIT: Nevermind, this number is now fed to different parameter in client which sets min count of streams, not actual requested count.
There was a problem hiding this comment.
Notice that the ReadSession creation takes two parameters - a maximal number of streams, and a preferred minimal number of streams, which the API tries to accommodate but treats it as a recommendation only, not binding limit. This logic is similar to what the Spark BigQuery connector is doing.
By the way, why not use the bigquery-connector-common library used by both Spark and Hive connectors?
There was a problem hiding this comment.
@davidrabinowitz that you for your input!
By the way, why not use the bigquery-connector-common library used by both Spark and Hive connectors?
I think this is a very important question, but I definitely am not in a position to answer it. Please allow me to gloss over it.
Notice that the ReadSession creation takes two parameters - a maximal number of streams, and a preferred minimal number of streams, which the API tries to accommodate but treats it as a recommendation only, not binding limit.
Indeed. And this PR switches us from setting the max, to setting the min.
The idea is that we actually don't want to have a static max. For example, if the data set is very large, we want to have a very large number of "reasonably sized" splits.
However, I don't know whether this is how it works.
@davidrabinowitz do we need to set both 'recommended min' and 'strict max' parameters? or can we just set the 'recommended min' ?
There was a problem hiding this comment.
Spark BigQuery connector seems to set both (https://github.com/search?q=repo%3AGoogleCloudDataproc%2Fspark-bigquery-connector+setPreferredMinStreamCount&type=code). They request at-least 3, and at most 20k by default (which is odd since IIRC 1k is the actual limit as documented at https://cloud.google.com/php/docs/reference/cloud-bigquery-storage/latest/V1.CreateReadSessionRequest#_Google_Cloud_BigQuery_Storage_V1_CreateReadSessionRequest__setPreferredMinStreamCount__).
Setting max might be useful for example if we want to avoid having too many splits.
There was a problem hiding this comment.
Sure, but what should be the max value? 10k?
if the actual limit is 1k, we're fine.
There was a problem hiding this comment.
By the way, why not use the bigquery-connector-common library used by both Spark and Hive connectors?
@davidrabinowitz @anoopj I sent DM in Trino community Slack. Let's continue the discussion there.
plugin/trino-bigquery/src/main/java/io/trino/plugin/bigquery/ReadSessionCreator.java
Outdated
Show resolved
Hide resolved
|
Please rebase @findepi |
Happy to, provided we have directional agreement about the change. |
|
Let's go ahead with the change. It makes sense to me. If we ever see that BigQuery gives errors when requesting a stream count in high concurrency workloads then we can introduce some config to adjust the "multiplier value" so that we don't run into quotas/limits. For now though, no changes requested. |
|
@findepi I'm in favor of having less configuration in connectors so 👍🏼 from me |
|
I don't see any requests for changes except editorial #22279 (comment). |
|
pls fix conflicts |
|
just rebased |
Similar to what we do in other connectors. There is no reason to create multiple splits just to return row count information.
No description provided.