Support Optimized Write #1198

sezruby · 2022-06-13T16:51:04Z

Description

Support OptimizeWrite described in https://docs.databricks.com/delta/optimizations/auto-optimize.html#how-optimized-writes-work

Fixes #1158

If OptimizeWrite is enabled, inject OptimizeWriteExchangeExec on top of the write plan and remove ShuffleExchangeExec or CoalesceExchange operation at the top of the plan to avoid unnecessary shuffle / stage.

In OptimizeWriteExchangeExec,

Repartition data
- RoundRobinPartitining for non partitioned data, HashPartitioning for partitioned data.
- Use spark.sql.shuffle.partitions for partitioning. We can introduce a new config like spark.sql.adaptive.coalescePartitions.initialPartitionNum if needed.
Rebalance partitions for write
- Step1 - merge small partitions (CoalescedPartitionSpec)
- Step2 - split large partitions (PartialReducerPartitionSpec)
- targetSize config: spark.databricks.delta.optimizeWrite.binSize (default: 128MB)

How to enable

Ref: https://docs.databricks.com/delta/optimizations/auto-optimize.html#enable-auto-optimize
We can enable OptimizeWrite using Spark session config or table property.

Spark session config
- spark.databricks.delta.optimizeWrite.enabled = true
- applied for write operations of all Delta tables)
Table property
- delta.autoOptimize.optimizeWrite = true

Spark session config is prior to the table property.

How was this patch tested?

Unit tests (+ more tests will be added)

Does this PR introduce any user-facing changes?

Yes, support OptimizeWrite

vkorukanti · 2022-06-14T17:50:36Z

Thank you @sezruby for creating this PR! This is a very useful feature. We are currently busy with the next release of Delta Lake. Will be reviewing the PR after the release.

scottsand-db · 2022-09-15T15:17:27Z

Can you please fix the conflicts?

Signed-off-by: Eunjin Song <sezruby@gmail.com>

tdas

Hello @sezruby my apologies for not being able to review this so long. i took a first pass and i think my biggest feedback is that i dont understand the overall optimization algorithm and the parameters. Typically for a feature like this we write design docs (see Optimize Zorder design doc) where we discuss the design choices and come to an agreement. Could you write a short doc explaining the algorithm? I want to understand the behavior and edge cases of this optimization.

tdas · 2022-09-08T22:06:22Z

core/src/main/scala/org/apache/spark/sql/delta/util/DeltaShufflePartitionsUtil.scala

+    })
+  }
+
+  private[sql] def removeTopRepartition(plan: SparkPlan): SparkPlan = {


Can you explain each of the cases with comments... these are pretty complicated on to reason about

Thanks for the review! I'll add some comments & classdoc and try to improve documentation in #1158

@tdas I added some example for OptimizeWrite partitioned data.
#1158 (comment)
Could you have a look and let me know if there is something unclear?

For non-partitioned data, here I use RoundRobinPartitioning but it could be inefficient in some cases as it distributes all rows into all partitions which is unnecessary. I think we could improve it later.

tdas · 2022-09-08T22:15:04Z

core/src/main/scala/org/apache/spark/sql/delta/util/DeltaShufflePartitionsUtil.scala

+  private[sql] def removeTopRepartition(plan: SparkPlan): SparkPlan = {
+    plan match {
+      case p@AdaptiveSparkPlanExec(inputPlan: ShuffleExchangeExec, _, _, _, _)
+        if !inputPlan.shuffleOrigin.equals(ENSURE_REQUIREMENTS) =>


I am trying to understand this and map it spark code. what you are trying to do is remove a shuffle if it wasnt added automatically by the planner to ensure requirement. Doesnt that mean if user asked for repartitioning by a certain way with an explicit programmatic API (e.g., DataFrame.repartition) we will be ignoring that completely?

Yes it's because for Optimize Write, we add repartition(partitionColumns)( + rebalancing) at top of the plan, so unnecessary repartition(n) or coalesce(n) could be removed.

#1158 - Things to do - 3 for detail

core/src/main/scala/org/apache/spark/sql/delta/util/DeltaShufflePartitionsUtil.scala

tdas · 2022-10-05T23:13:08Z

core/src/main/scala/org/apache/spark/sql/execution/OptimizeWriteExchangeExec.scala

+import scala.concurrent.Future
+import scala.concurrent.duration.Duration
+
+case class OptimizeWriteExchangeExec(


Class docs are needed with a full explanation of how the optimization occurs, what the algorithm is like, what are the parameters. Its really hard to understand the over all algorithm from the code without an overview here.

Signed-off-by: Eunjin Song <sezruby@gmail.com>

Kimahriman

Would be great to get some momentum going on this again. Seems like it makes sense after understanding how Spark does rebalancing. My question would be if it works for streaming writes, and if so might be good to add a test for that?

Kimahriman · 2023-01-23T12:28:34Z

core/src/main/scala/org/apache/spark/sql/delta/util/DeltaShufflePartitionsUtil.scala

+        } else {
+          mapStartIndices(i + 1)
+        }
+        val dataSize = startMapIndex.until(endMapIndex).map(mapPartitionSizes(_)).sum


There was an update on the Spark side to make this more performant: apache/spark@9e1d00c

Is the main reason to not just use the Spark versions directly to add the configurable mergedPartitionFactor, or to not rely on those internal helper functions?

Yes it's for both; make configurable and not rely on spark internal util which can change frequently. But I'm also okay to use Spark one. I'm just waiting for databricks team to get back on this. (and auto compaction too)

https://learn.microsoft.com/en-us/azure/synapse-analytics/spark/optimize-write-for-apache-spark

FYI, the code is being used in prod at least 6 months and no major issue so far.
We might need to increase binSize config / PARQUET COMPRESSION RATIO for larger files. (60~80mb avg for now)

Hmm I tried building and using this myself and I don't seem to be getting my large partitions split, gonna add some more logging to try to see why/what's happening.

Yes it's the case this approach (Spark rebalance logic) cannot handle. Because the unit of rebalancing is determined by the source partition layout. If the data is skewed or only few number of partitions, it cannot be rebalanced properly.
Please refer the figure in #1158 (comment)

If the source dataframe consists of all same key=1 and 10GB of 1 partition, it cannot be split.
e.g. df.repartition(1, col("key")).select(col("key")).write.format("delta").parquet("path")

I removed redundant repartition execution plan on top of the child plan, but it cannot cover all the cases.

The smallest unit of splitting is the single task map output for a single reducer ID right? That wasn't what I was seeing, where I had all my map tasks shuffle write < 1 GB, but I had some reducer tasks reading > 10 GB of shuffle data. After digging through how map output sizes work a little bit, I'm gonna try again and see if this is some weird side effect of HighlyCompressedMapStatus for large numbers of reducing partitions (> 2000 by default). Only thought is some weird effect of a lot of small blocks being "averaged out" to compute map output sizes per reducer.

That appeared to be it. I dropped my shuffle partitions to 1k and it behaved as I would expect. Not sure how common my case would be with particular types of data skew, but maybe it would be good to log a warning if your shuffle partitions exceeds the threshold for using HighlyCompressedMapStatus, it can limit the ability to properly split skewed partitions, because average map output size is used for individual map reducer outputs, instead of the real value.

Kimahriman · 2023-04-09T14:03:30Z

Gentle ping on this again, just started using this in our production environment and would be great not to have to maintain my own Delta fork 😅

sezruby · 2023-04-10T17:15:55Z

Please let me know if someone is ready to review. I'll rebase the PR then.

isunli · 2023-05-16T16:43:55Z

any update on this PR?

delta-io#1198

cb-sukumarnataraj · 2023-09-01T13:53:10Z

Any update on this PR?.

sezruby · 2023-09-06T16:57:15Z

Hi @tdas @scottsand-db Any update? So no plan to deliver this feature to OSS delta?

adityakumar84 · 2023-09-22T07:54:55Z

I am using your PR for optimize write in my EMR streaming use cases.

I want that optimize write should create around 128m parquet files but it is topping at 64m.

Configuration I am using is as below
"spark.sql.catalog.spark_catalog": "org.apache.spark.sql.delta.catalog.DeltaCatalog",
"spark.databricks.delta.autoCompact.enabled": "true",
"spark.databricks.delta.optimizeWrite.enabled": "true",
"spark.databricks.delta.autoCompact.minNumFiles": "500",
"spark.sql.adaptive.enabled": "true",
"spark.sql.adaptive.coalescePartitions.enabled": "true",
"spark.sql.adaptive.advisoryPartitionSizeInBytes": "128m",
"spark.sql.adaptive.coalescePartitions.minPartitionNum": "1",
"spark.sql.adaptive.coalescePartitions.initialPartitionNum": "200",
"spark.databricks.delta.optimizeWrite.binSize": "134217728

2023-09-22 08:22:54 61.9 MiB part-00000-eb8c145c-7ac4-420c-8b85-c6fd58861e40.c000.snappy.parquet
2023-09-22 08:24:55 43.7 MiB part-00001-4fbd8215-5129-48aa-b858-d2ab9d2ca7a6.c000.snappy.parquet
2023-09-22 08:24:58 54.8 MiB part-00000-de4d1a37-543c-4c80-a2dd-9cbc06af7b3c.c000.snappy.parquet
2023-09-22 08:27:06 52.7 MiB part-00000-c6cf2c15-8dc3-4346-a03c-eb83171fdf46.c000.snappy.parquet
2023-09-22 08:27:06 59.3 MiB part-00001-a5f3db2e-e352-45b0-aa1b-bbf1570ebb6f.c000.snappy.parquet
2023-09-22 08:28:56 69.1 MiB part-00000-697cca77-e7ea-4267-9f7b-3fd406678b31.c000.snappy.parquet
2023-09-22 08:31:03 52.8 MiB part-00000-3dfbf27e-53f3-4f35-8955-7675d0ec86bf.c000.snappy.parquet
2023-09-22 08:31:05 60.5 MiB part-00001-6b7bc1d7-9882-4ba0-a01b-619a0a424ae4.c000.snappy.parquet
2023-09-22 08:32:58 32.5 MiB part-00001-ea4d882a-f182-48d4-a3e8-bc23260e617e.c000.snappy.parquet
2023-09-22 08:33:03 54.1 MiB part-00000-4c53bf10-eede-45c3-b178-f9d4f2895956.c000.snappy.parquet
2023-09-22 08:35:00 45.7 MiB part-00001-138bdaab-8d77-4faf-af49-489c4cfad4f6.c000.snappy.parquet
2023-09-22 08:35:01 51.7 MiB part-00000-62a2bdfa-9dea-41ff-9f8c-4e3ebcb12d04.c000.snappy.parquet
2023-09-22 11:42:10 52.0 MiB part-00000-64bccc7b-2664-4e67-b478-413a5a7554db.c000.snappy.parquet
2023-09-22 11:42:12 57.9 MiB part-00001-b79e83d9-c65e-438b-a555-aa2028320ea5.c000.snappy.parquet
2023-09-22 11:45:20 56.4 MiB part-00001-1ef411fb-160c-4cba-a5e7-cdee540295a5.c000.snappy.parquet
2023-09-22 11:45:20 57.0 MiB part-00004-fa51fd84-c4d3-4980-9a88-5566283d3234.c000.snappy.parquet
2023-09-22 11:45:20 57.5 MiB part-00005-80e03284-a8a6-4937-857c-1c434527c956.c000.snappy.parquet
2023-09-22 11:45:20 57.8 MiB part-00000-a347b170-d15f-4328-b402-ccdc099e94da.c000.snappy.parquet
2023-09-22 11:45:20 57.9 MiB part-00006-d566e257-f0cf-4cf9-b968-cad449784226.c000.snappy.parquet
2023-09-22 11:45:20 57.9 MiB part-00009-18a54bcf-3d53-4766-b485-62332266c974.c000.snappy.parquet
2023-09-22 11:45:20 58.5 MiB part-00003-8cf66203-9a32-41b5-a617-a3a9e9b2e465.c000.snappy.parquet
2023-09-22 11:45:21 38.7 MiB part-00010-1b8e65ae-b366-46fa-818e-d8ac95fd5904.c000.snappy.parquet
2023-09-22 11:45:22 57.1 MiB part-00007-05dc33ff-5125-435f-b0f1-f4215fba9d5b.c000.snappy.parquet
2023-09-22 11:45:22 58.9 MiB part-00008-5c99002a-1d7b-4db5-8c14-8e3f966773ec.c000.snappy.parquet
2023-09-22 11:45:23 57.6 MiB part-00002-6844ca06-b437-486f-b15b-ebf93fc0e740.c000.snappy.parquet

Any suggestions?

sezruby · 2023-09-26T01:15:52Z

@adityakumar84 you can try to increase the binSize config. It's used for row format size in memory.
The result parquet size, we cannot control it precisely as it depends on data property / parquet compression ratio.
To make more accurate, we need to collect parquet compression ratio history from previous wrtie job... etc

scottsand-db requested a review from vkorukanti June 13, 2022 18:10

sezruby force-pushed the optimizewrite branch from 82fbb8a to cb31280 Compare June 15, 2022 17:04

sezruby force-pushed the optimizewrite branch from cb31280 to 8aefaca Compare July 19, 2022 00:02

dennyglee mentioned this pull request Sep 14, 2022

Roadmap 2022 H2 (discussion) #1307

Open

scottsand-db requested review from tdas and removed request for vkorukanti September 14, 2022 18:15

Support Optimized Write

72e6555

Signed-off-by: Eunjin Song <sezruby@gmail.com>

sezruby force-pushed the optimizewrite branch from 8aefaca to 72e6555 Compare September 15, 2022 22:44

tdas requested changes Oct 5, 2022

View reviewed changes

sezruby force-pushed the optimizewrite branch 2 times, most recently from 13927b5 to a2cb8b6 Compare October 18, 2022 00:25

Add comment

e2d12dc

Signed-off-by: Eunjin Song <sezruby@gmail.com>

sezruby force-pushed the optimizewrite branch from a2cb8b6 to e2d12dc Compare October 18, 2022 00:40

sezruby requested a review from tdas October 24, 2022 22:19

Kimahriman reviewed Jan 23, 2023

View reviewed changes

keen85 mentioned this pull request Feb 5, 2023

[Feature Request] control the number of output/file size if repartitionBeforeWrite is enabled #1585

Open

3 tasks

jlowe mentioned this pull request Feb 13, 2023

Support Delta Lake optimized write on Databricks NVIDIA/spark-rapids#7749

Merged

shenavaa added a commit to shenavaa/delta that referenced this pull request Jul 21, 2023

Support Optimized Write

9956a61

delta-io#1198

felipepessoto mentioned this pull request Oct 6, 2023

[Spark] Implement optimized write. #2145

Closed

5 tasks

rasidhan mentioned this pull request Oct 10, 2023

[Feature Request] OPTIMIZED WRITE #1158

Closed

1 task

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support Optimized Write #1198

Support Optimized Write #1198

sezruby commented Jun 13, 2022 •

edited

Loading

vkorukanti commented Jun 14, 2022

scottsand-db commented Sep 15, 2022

tdas left a comment

tdas Sep 8, 2022

sezruby Oct 10, 2022

sezruby Oct 12, 2022

tdas Sep 8, 2022

sezruby Oct 10, 2022

tdas Oct 5, 2022

Kimahriman left a comment

Kimahriman Jan 23, 2023

sezruby Jan 26, 2023

sezruby Jan 26, 2023

Kimahriman Jan 30, 2023

sezruby Jan 31, 2023

Kimahriman Jan 31, 2023

Kimahriman Jan 31, 2023

Kimahriman commented Apr 9, 2023

sezruby commented Apr 10, 2023

isunli commented May 16, 2023

cb-sukumarnataraj commented Sep 1, 2023

sezruby commented Sep 6, 2023

adityakumar84 commented Sep 22, 2023

sezruby commented Sep 26, 2023

Support Optimized Write #1198

Are you sure you want to change the base?

Support Optimized Write #1198

Conversation

sezruby commented Jun 13, 2022 • edited Loading

Description

How to enable

How was this patch tested?

Does this PR introduce any user-facing changes?

vkorukanti commented Jun 14, 2022

scottsand-db commented Sep 15, 2022

tdas left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Kimahriman left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Kimahriman commented Apr 9, 2023

sezruby commented Apr 10, 2023

isunli commented May 16, 2023

cb-sukumarnataraj commented Sep 1, 2023

sezruby commented Sep 6, 2023

adityakumar84 commented Sep 22, 2023

sezruby commented Sep 26, 2023

sezruby commented Jun 13, 2022 •

edited

Loading