Very slow performance

[rajkve]

Hi All,

Config - Databricks 6.6 (Spark 2.45)
Target - Azure SQL DB Premium P4

This connector , code

FLOC_VW.write \ .format("com.microsoft.sqlserver.jdbc.spark") \ .mode("overwrite") \ .option("url", url) \ .option("dbtable", tableName) \ .option("user", Username) \ .option("password", Password) \ .option("BEST_EFFORT","true") \ .option("tabLock","true") \ .save()

Minutes taken - 53 on Average

Same configuration but - spark connector - Performance 4x faster <10 mins of average

Any suggestions where it could be going wrong?

[shivsood]

[old spark connector](old spark connector and this should have the same perf as both use the bulk APIs. ) and mssql spark connector should have the same perf as both use the bulk APIs. I have not benchmarked this against each other though.

Can u please share configuration details e.g. data set used, spark configuration, time in seconds to write by each, methodology etc.

[rajkve]

Hi -

Spark cluster - 2.4.5, Scala 11
Driver - 32 GB memory , 16 cores
Worker - 23 GB 4 Cores (Min nodes 5, max nodes 20)

Source - ADLS GEN1 Parquet file size - 500 MB (5 Million records)

Target - Azure SQL DB (Premium P4)

Table size - 12 GB (Most of the fields are text fields)

Data upload performance -

using this connector - 55 mins (BEST_EFFORT + TAB_LOCK = true) - source code in the first post above
using Spark connector - 10 Mins

[sebbm]

Hi, maybe you should try to increase the batchSize as this is set to 1.000 by default.

[rajkve]

Hi, I tried with batch size = 10000, 50000, 500000 - it takes approximately the same time.

[denzilribeiro]

@rajkve is your base table a heap or have columnstore index?

[sirurza]

We have the same problem, parquet in ADLS GEN2
To table with 30 columns, with their types; char, varchar, int, bit and decimal (18,6).
The table does not have any contraint, primary keys, or indexes.

Method	Nº Rows	BatchSize	TableLock	Timeout	ReliabilityLevel	Time (min)	Time stimate (min)
bulkCopyToSqlDB	1.000.000	1.000	true	1000		1,65	36,3
bulkCopyToSqlDB	1.000.000	500.000	true	1000		1,59	34,98
bulkCopyToSqlDB	5.000.000	1.000	true	1000		8,11	35,684
bulkCopyToSqlDB	5.000.000	500.000	true	1000		7,6	33,44
bulkCopyToSqlDB	10.000.000	1.000	true	1000		15,54	34,188
bulkCopyToSqlDB	10.000.000	100.000	true	1000		15,49	34,078
JDBC	1.000.000	1.000				2,18	47,96
JDBC	1.000.000	500.000				2,77	60,94
JDBC	5.000.000	1.000				10,66	46,904
JDBC	5.000.000	500.000				13,68	60,192
JDBC	10.000.000	1.000				21,1	46,42
JDBC	10.000.000	10.000				21,1	46,42
jdbc.spark	1.000.000	1.000	true		BEST_EFFORT	2,21	48,62
jdbc.spark	1.000.000	10.000	true		BEST_EFFORT	2,12	46,64
jdbc.spark	1.000.000	100.000	true		BEST_EFFORT	2,12	46,64
jdbc.spark	1.000.000	500.000	true		BEST_EFFORT	2,12	46,64
jdbc.spark	5.000.000	1.000	true		BEST_EFFORT	10,63	46,772

[shivsood]

@sirurza I am not able to interpret these results.

• how do interpret time v/s Time stimate?
• Does 2,21 represent 2 minutes 21 seconds.
• What connectors do bulkCopyToSQLDB, JDBC and JDBC.spark represent?

[sirurza]

Hi
Our table have 22000000 rows, stimate is the total time in completed task.
The time column is the time it takes to complete the No.Rows

The values are shown in decimals, that is, the decimal part is the seconds, in other words 36.3 is 36 min and the seconds are 3 * 60/100, the result is 36min with 11 seconds.

bulkCopyToSQLDB = https://github.com/Azure/azure-sqldb-spark

[DataBanshee]

Hi Team,

I'm facing a similar issue, the default connector is much faster then the sql-spark-connector even though I tinkered with the Table Lock and the Batch Size... :(
I was hoping that this connector could help me out in writing 70M rows with 8 nodes (28GB of Ram each)

[denzilribeiro]

@sirurza / @DataBanshee can you please post your snippet of the Dataframe loading the data?
During the run what are predominant waits?
a. Current waits - https://www.sqlskills.com/blogs/paul/updated-sys-dm_os_waiting_tasks-script-2/
b. Take 2 snapshots say 30 seconds apart say - https://www.sqlskills.com/blogs/paul/capturing-wait-statistics-period-time/

[DataBanshee]

Hey @denzilribeiro
Thank you for you help :)
The job seems to get stuck in the last task and stay hanging. 1M rows is the maximum I could achieve without a frozen task

I also tried to size up the cluster on the Spark side with no luck... I have been on this for days and still can't understand why Millions of rows are such challenge
Grateful for your help !

[denzilribeiro]

@DataBanshee - Couple things I am seeing
a. If using a heap use tableLock "true"
b. For a single task - are you reading from ADLS/HDFS that is partitioned or something where 1 task gets more rows or are they evenly distributed?

Will be easier to take this on email -- email me at denzilr at microsoft .

[ankitbko]

@denzilribeiro Can confirm this connector bulk insert slower than the older Spark Connector. I have TPC-DS dataset in parquet and a notebook containing code to insert into Azure SQL hyperscale using both old and new connector. The old connector is atleast 2x faster. I was writing a blog about the new connector when I stumbled across this issue and can include the perf comparison there which will take few days to complete. Or I can show you right away if you can ping me at anksinha

[ankitbko]

I ran a benchmark comparing timing by inserting 50 million records in databases with different indexing strategy. You can find it here. Scroll down to the bottom so see comparison between old and new connector.