Databricks Community

NOOR_BASHASHAIK · ‎10-22-2022

Dears

I was trying to check what Azure Databricks VM type is best suited for executing OPTIMIZE with ZORDER on a single timestamp value (but string data type) column for around 5000+ tables in the Delta Lake.

I chose Standard_F16s_v2 with 6 workers & 1 driver. (> so, this gives total 112 CPUs, 185 GB Memory, 7571.4 GB OS disk storage). We are in 9.1 LTS. Given below is some information on how the CPU and the memory usage looks like. Also, in "Storage" tab of Spark UI, I do not see any Cache misses but there is lot of shuffle read in "Executors" tab. The job is running for 7+ hours, and I see still 216 active stages, and 219 pending stages (this number keeps changing). spark.sql.shuffle.partitions is set to default.

My question is : based on the information provided here, which series VM is best suitable? Also, any other sort of optimization that can be done? If I go with the best practice to set shuffle partitions value as 2 * number of cores, the default value still suits.

Hubert-Dudek · ‎10-23-2022

Your chart; its nice (compared to other ones which I see); it seems you utilize your cluster correctly.

You can check in Spark UI for data spills - are Spark partitions fit in memory?

Any cluster should be okay if your disk partition size for your tables is around 200 MB (optimal size). Then, of course, you can do benchmarking. Usually, new versions of machines are a bit faster.

Tooptimizationstimzations for tables where you append data is good to use disk partitioning per date or month. Then, do OPTIMIZATION with the WHERE clause to limit it to only new partitions.

NOOR_BASHASHAIK · ‎10-24-2022

Hi Hubert,

In screenshot#3 I put above, there is plenty of shuffle read? Is that what needs to be checked for data spills?

Also, please let me know how I can set this flag - disk partition size for your tables is around 200 MB (optimal size).

Appreciate, and thanks in advance.

jose_gonzalez · ‎10-24-2022

Hi,

The Standard_F16s_v2 is a compute optimize type machine.

On the other-hand, for delta optimize (both bin-packing and Z-Ordering), we recommend Stabdard_DS_v2-series. Also, follow Hubert's recommendations.

NOOR_BASHASHAIK · ‎10-24-2022

Hi Jose,

Thanks for the quick reply. Yes, I plan to test now with DdS_V5 series, as DdS series seems to be delta cache accelerated where as other DS series seem to be falling under "General Purpose" workload machines.