Potential bottleneck Azure Databricks Cluster - azure

I am facing an issue with dabricks cluster. When saving a 20 billon rows dataset, it is taking days. But when I analyse cpu performance, it comes up that the following usage:
-46.8% iddle
-9.2% user
-43.4% wait
-0.7 system.
The cluster is a multi node with 4 workers and photon aceleration technology. Memory performance has 100gb free from a total of 700gb.
What does "wait" means? Is there something i can do to speed up the process? What else can i analyze to improve times?
Best regards

Related

GC and shuffle read is high (red) in databricks cluster, how to tune this?

I need recommendation for cluster design in databricks, we have a ETL batch load running every 20 minutes.
there are 25+ notebooks doing straight merge in silver layer tables( fact/dimensions)
cluster config are as follows:
instance type- F64- compute optimised
worker nodes -3 - 128 gb memroy X 64 cores
driver node - 1 memory optimized - 64 GB X8 cores
we need to run minimize execution time and increase parallelism
I am attaching snapshot from SPARKUI of databricks cluster (executor page) for your refrence.
[1]: https://i.stack.imgur.com/qMFyf.png
I see red flag for GC time and shuffle read, GC time turns out more than 10% of total task time.
How can we bring this down and we missing our SLA for load cycle.
Thanks
Try increasing the executors and nodes and also giving more memory per executor. Also, you can see which line in your code is specifically taking long by looking at logs.
You need to ensure that you are not performing actions that involve a lot of data shuffling.

decide no of partition in spark (running on YARN) based on executer ,cores and memory

How to decide no of partition in spark (running on YARN) based on executer, cores and memory.
As i am new to spark so doesn't have much hands on real scenario
I know many things to consider to decide the partition but still any production general scenario explanation in detail will be very helpful.
Thanks in advance
One important parameter for parallel collections is the number of
partitions to cut the dataset into. Spark will run one task for each
partition of the cluster. Typically you want 2-4 partitions for each
CPU in your cluster
the number of parition is recommended to be 2/4 * the number of cores.
so if you have 7 executor with 5 core , you can repartition between 7*5*2 = 70 and 7*5*4 = 140 partition
https://spark.apache.org/docs/latest/rdd-programming-guide.html
IMO with spark 3.0 and AWS EMR 2.4.x with adaptive query execution you're often better off letting spark handle it. If you do want to hand tune it the answer can often times be complicated. One good option is to have 2 or 4 times the number of cpus available. While this is useful for most datasizes it becomes problematic with very large and very small datasets. In those cases it's useful to aim for ~128MB per partition.

How to speedup node joining process in cassandra cluster

I have a cluster 4 cassandra nodes. I have recently added a new node but data processing is taking too long. Is there a way to make this process faster ? output of nodetool
Less data per node. Your screenshot shows 80TB per node, which is insanely high.
The recommendation is 1TB per node, 2TB at most. The logic behind this is bootstrap times get too high (as you have noticed). A good Cassandra ring should be able to rapidly recover from node failure. What happens if other nodes fail while the first one is rebuilding?
Keep in mind that the typical model for Cassandra is lots of smaller nodes, in contrast to SQL where you would have a few really powerful servers. (Scale out vs scale up)
So, I would fix the problem by growing your cluster to have 10X - 20X the number of nodes.
https://groups.google.com/forum/m/#!topic/nosql-databases/FpcSJcN9Opw

Processing Pipeline using Spark SQL- jobs, stages and DAG sizes

I have a processing pipeline that is built using Spark SQL. The objective is to read data from Hive in the first step and apply a series of functional operations (using Spark SQL) in order to achieve the functional output. Now, these operations are quite in number (more than 100), which means I am running around 50 to 60 spark sql queries in a single pipeline. While the application completes successfully without any issues, my focus area has shifted to optimizing the overall process. I have been able to speed up the executions using spark.sql.shuffle.partitions, changing the executor memory and reducing the size of the spark.memory.fraction from default 0.6 to 0.2. I got great benefits by doing all these changes and the over all execution time reduced from 20-25 mins to around 10 mins. Data volume is around 100k rows (source side).
The observations that I have from the Cluster are:
-The number of jobs triggered as apart of application id are 235.
-The total number of stages across all the jobs created are around 600.
-8 executors are used in a two node cluster (64 GB RAM in total with 10 cores).
-The resource manager UI of Yarn (for an application id) becomes very slow to retrieve the details of jobs/stages.
In one of the videos of Spark tuning, I heard that we should try to reduce the number of stages to a bare minimum, also DAG size should be smaller. What are the guidelines to do this. How to find the number of shuffles that are happening (my SQLs have many joins and group by clauses).
I would like to have suggestions on the above scenario of what possible things I can do in order to improvise the performance and handle the data skews in the SQL queries that are JOIN/GROUP_BY heavy.
Thanks

Does Spark incur the same amount of overhead as Hadoop for vnodes?

I just read https://stackoverflow.com/a/19974621/260805. Does Spark (specifically Datastax's Cassandra Spark connector) incur the same amount of overhead as Hadoop when reading from a Cassandra cluster? I know Spark uses threads more heavily than Hadoop does.
Performance with vnodes and without in the connector should be basically the same. With hadoop each vnode split generated it's own task which created a large amount of overhead.
With Spark, tasks contain the token ranges from multiple vnodes and are merged into a single task and the overall task overhead is lower. There is a slight locality issue where it becomes difficult to get balanced numbers of tasks for all the nodes in the C* cluster with smaller data sizes. This issue is being worked on in SPARKC-43.
I'll give three separate answers. I apologize for the rather unstructured answer, but it's been building up over time:
A previous answer:
Here's one potential answer: Why not enable virtual node in an Hadoop node?. I quote:
Does this also apply to Spark?
No, if you're using the official DataStax spark-cassandra-connector. It can process multiple token ranges in a single Spark task. There is still some minor performance hit, but not as huge as with Hadoop.
A production benchmark
We ran a Spark job against a vnode-enabled Cassandra (Datastax Enterprise) datacenter with 3 nodes. The job took 9.7 hours. Running the same job on for slightly less data, using 5 non-vnode nodes, a couple of weeks back took 8.8 hours.
A controlled benchmark
To further test the overhead we ran a controlled benchmark on a Datastax Enterprise node in a single-node cluster. For both vnode enabled/disabled the node was 1) reset, 2) X number of rows were written and then 3) SELECT COUNT(*) FROM emp in Shark was executed a couple of times to get a cold vs. hot cache times. X tested were 10^{0-8}.
Assuming that Shark is not dealing with vnodes in any way, the average (quite stable) overhead for vnodes were ~28 seconds for cold Shark query executions and 17 seconds for hot executions. The latency difference did generally not vary with data size.
All the numbers for the benchmark can be found here. All scripts used to run the benchmark (see output.txt for usage) can be found here.
My only guess why there was a difference between "Cold diff" and "Hot diff" (see spreadsheet) is that it took Shark some time to create metadata, but this is simply speculation.
Conclusion
Our conclusion is that the overhead of vnodes is a constant time between 13 and 30 seconds, independent of data size.

Resources