Looking for efficient partitioning strategies for my dataframe when storing my dataframe in the delta table.
My current dataframe 1.5000.000 rowa it takes 3.5h to move data from dataframe to delta table.
Looking for a more efficient way to do this writing I decided to try different columns of my table as partitioning columns.I searched for the cardinality of my columns and selected the following ones.
column1 = have 3 distinct_values
column2 = have 7 distinct values
column3 = have 26 disctinc values
column4 = have 73 distinc values
column5 = have 143 distinc values
column6 = have 246 distinct values
column7 = have 543 disctinc values
cluster: 64GB, 8 cores
using the folloging code in my notebook
df.write.partitionBy("column_1").format("delta").mode("overwrite").save(partition_1)
..
df.write.partitionBy("column_7").format("delta").mode("overwrite").save(partition7)
Thus, I wanted to see which partitioning strategy would bring better results: a column with high cardinality, one with low cardinality or one in between.
To my surprise this has not had any effect as it has taken practically the same time in all of them with differences of a few minutes but all of them with + 3h.
why have I failed ? is there no advantage to partitioning ?
When you use Delta (either Databricks or OSS Delta 1.2.x, better 2.0) then often you may not need to use partitioning at all for following reasons (that aren't applicable for Parquet or other file formats):
Delta supports data skipping that allows to read only necessary files, especially effective when you use it in combination with OPTIMIZE ZORDER BY that will put related data closer to each other.
Bloom filters allow to skip files even more granularly.
The rules of thumb of using partitioning with Delta lake tables are following:
use it when it will benefit queries, especially when you perform MERGE into the table, because it allows to avoid conflicts between parallel transactions
when it helps to delete old data (for example partitioning by date)
when it really benefits your queries. For example, you have data per country, and most of queries will use country as a part of condition. Or for example, when you partition by date, and querying data based on the time...
In all cases, don't use partitioning for high cardinality columns (hundreds of values) and having too many partition columns because in most cases it lead to creation of small files that are less efficient to read (each file is accessed separately), plus it leads to increased load to the driver as it needs to keep metadata for each of the file.
Related
I have a case where i am trying to write some results using dataframe write into S3 using the below query with input_table_1 size is 13 Gb and input_table_2 as 1 Mb
input_table_1 has columns account, membership and
input_table_2 has columns role, id , membership_id, quantity, start_date
SELECT
/*+ BROADCASTJOIN(input_table_2) */
account,
role,
id,
quantity,
cast(start_date AS string) AS start_date
FROM
input_table_1
INNER JOIN
input_table_2
ON array_contains(input_table_1.membership, input_table_2.membership_id)
where membership array contains list of member_ids
This dataset write using Spark dataframe is generating around 1.1TiB of data in S3 with around 700 billion records.
We identified that there are duplicates and used dataframe.distinct.write.parquet("s3path") to remove the duplicates . The record count is reduced to almost 1/3rd of the previous total count with around 200 billion rows but we observed that the output size in S3 is now 17.2 TiB .
I am very confused how this can happen.
I have used the following spark conf settings
spark.sql.shuffle.partitions=20000
I have tried to do a coalesce and write to s3 but it did not work.
Please suggest if this is expected and when can be done ?
There's two sides to this:
1) Physical translation of distinct in Spark
The Spark catalyst optimiser turns a distinct operation into an aggregation by means of the ReplaceDeduplicateWithAggregate rule (Note: in the execution plan distinct is named Deduplicate).
This basically means df.distinct() on all columns is translated into a groupBy on all columns with an empty aggregation:
df.groupBy(df.columns:_*).agg(Map.empty).
Spark uses a HashPartitioner when shuffling data for a groupBy on respective columns. Since the groupBy clause in your case contains all columns (well, implicitly, but it does), you're more or less randomly shuffling data to different nodes in the cluster.
Increasing spark.sql.shuffle.partitions in this case is not going to help.
Now on to the 2nd side, why does this affect the size of your parquet files so much?
2) Compression in parquet files
Parquet is a columnar format, will say your data is organised in columns rather than row by row. This allows for powerful compression if data is adequately laid-out & ordered. E.g. if a column contains the same value for a number of consecutive rows, it is enough to write that value just once and make a note of the number of repetitions (a strategy called run length encoding). But Parquet also uses various other compression strategies.
Unfortunately, data ends up pretty randomly in your case after shuffling to remove duplicates. The original partitioning of input_table_1 was much better fitted.
Solutions
There's no single answer how to solve this, but here's a few pointers I'd suggest doing next:
What's causing the duplicates? Could these be removed upstream? Or is there a problem with the join condition causing duplicates?
A simple solution is to just repartition the dataset after distinct to match the partitioning of your input data. Adding a secondary sorting (sortWithinPartition) is likely going to give you even better compression. However, this comes at the cost of an additional shuffle!
As #matt-andruff pointed out below, you can also achieve this in SQL using cluster by. Obviously, that also requires you to move the distinct keyword into your SQL statement.
Write your own deduplication algorithm as Spark Aggregator and group / shuffle the data just once in a meaningful way.
I am trying to understand the performance impact on the partitioning scheme when Spark is used to query a hive table. As an example:
Table 1 has 3 partition columns, and data is stored in paths like
year=2021/month=01/day=01/...data...
Table 2 has 1 partition column
date=20210101/...data...
Anecdotally I have found that queries on the second type of table are faster, but I don't know why, and I don't why. I'd like to understand this so I know how to design the partitioning of larger tables that could have more partitions.
Queries being tested:
select * from table limit 1
I realize this won't benefit from any kind of query pruning.
The above is meant as an example query to demonstrate what I am trying to understand. But in case details are important
This is using s3 not HDFS
The data in the table is very small, and there are not a large number of partitons
The time for running the query on the first table is ~2 minutes, and ~10 seconds on the second
Data is stored as parquet
Except all other factors which you did not mention: storage type, configuration, cluster capacity, the number of files in each case, your partitioning schema does not correspond to the use-case.
Partitioning schema should be chosen based on how the data will be selected or how the data will be written or both. In your case partitioning by year, month, day separately is over-partitioning. Partitions in Hive are hierarchical folders and all of them should be traversed (even if using metadata only) to determine the data path, in case of single date partition, only one directory level is being read. Two additional folders: year+month+day instead of date do not help with partition pruning because all columns are related and used together always in the where.
Also, partition pruning probably does not work at all with 3 partition columns and predicate like this: where date = concat(year, month, day)
Use EXPLAIN and check it and compare with predicate like this where year='some year' and month='some month' and day='some day'
If you have one more column in the WHERE clause in the most of your queries, say category, which does not correlate with date and the data is big, then additional partition by it makes sense, you will benefit from partition pruning then.
Running databricks to read csv files and then saving as a partitioned delta table.
Total records in file are 179619219 . It is being split on COL A (8419 unique values) and Year ( 10 Years) and Month.
df.write.partitionBy("A","year","month").format("delta") \
.mode("append").save(path)
Job gets stuck on the write step and aborts after running for 5-6 hours
This is very bad partitioning schema. You simply have too many unique values for column A, and additional partitioning is creating even more partitions. Spark will need to create at least 90k partitions, and this will require creation a separate files (small), etc. And small files are harming the performance.
For non-Delta tables, partitioning is primarily used to perform data skipping when reading data. But for Delta lake tables, partitioning may not be so important, as Delta on Databricks includes things like data skipping, you can apply ZOrder, etc.
I would recommend to use different partitioning schema, for example, year + month only, and do OPTIMIZE with ZOrder on A column after the data is written. This will lead to creation of only few partitions with bigger files.
I have my data in a delta lake in ADLS and am reading it through Databricks. The data is partitioned by year and date and z ordered by storeIdNum, where there are about 10 store Id #s, each with a few million rows per date. When I read it, sometimes I am reading one date partition (~20 million rows) and sometimes I am reading in a whole month or year of data to do a batch operation. I have a 2nd much smaller table with around 75,000 rows per date that is also z ordered by storeIdNum and most of my operations involve joining the larger table of data to the smaller table on the storeIdNum (and some various other fields - like a time window, the smaller table is a roll up by hour and the other table has data points every second). When I read the tables in, I join them and do a bunch of operations (group by, window by and partition by with lag/lead/avg/dense_rank functions, etc.).
My question is: should I have the date in all of the joins, group by and partition by statements? Whenever I am reading one date of data, I always have the year and the date in the statement that reads the data as I know I only want to read from a certain partition (or a year of partitions), but is it important to also reference the partition col. in windows and group bus for efficiencies, or is this redundant? After the analysis/transformations, I am not going to overwrite/modify the data I am reading in, but instead write to a new table (likely partitioned on the same columns), in case that is a factor.
For example:
dfBig = spark.sql("SELECT YEAR, DATE, STORE_ID_NUM, UNIX_TS, BARCODE, CUSTNUM, .... FROM STORE_DATA_SECONDS WHERE YEAR = 2020 and DATE='2020-11-12'")
dfSmall = spark.sql("SELECT YEAR, DATE, STORE_ID_NUM, TS_HR, CUSTNUM, .... FROM STORE_DATA_HRS WHERE YEAR = 2020 and DATE='2020-11-12'")
Now, if I join them, do I want to include YEAR and DATE in the join, or should I just join on STORE_ID_NUM (and then any of the timestamp fields/customer Id number fields I need to join on)? I definitely need STORE_ID_NUM, but I can forego YEAR AND DATE if it is just adding another column and makes it more inefficient because it is more things to join on. I don't know how exactly it works, so I wanted to check as by foregoing the join, maybe I am making it more inefficient as I am not utilizing the partitions when doing the operations? Thank you!
The key with delta is to choose the partitioned columns very well, this could take some trial and error, if you want to optimize the performance of the response, a technique I learned was to choose a filter column with low cardinality (you know if the problem is of time series, it will be the date, on the other hand if it is about a report for all clients in that case it may be convenient to choose your city), remember that if you work with delta each partition represents a level of the file structure where its cardinality will be the number of directories.
In your case I find it good to partition by YEAR, but I would add the MONTH given the number of records that would help somewhat with the dynamic pruning of spark
Another thing you can try is to use BRADCAST JOIN if the table is very small compared to the other.
Broadcast Hash Join en Spark (ES)
Join Strategy Hints for SQL Queries
The latter link explains how dynamic pruning helps in MERGE operations.
How to improve performance of Delta Lake MERGE INTO queries using partition pruning
In my case I have a table structure like this:
table_1 {
entity_uuid text
,fk1_uuid text
,fk2_uuid text
,int_timestamp bigint
,cnt counter
,primary key (entity_uuid, fk1_uuid, fk2_uuid, int_timestamp)
}
The text columns are made up of random strings. However, only entity_uuid is truly random and evenly distributed. fk1_uuid and fk2_uuid have much lower cardinality and may be sparse (sometimes fk1_uuid=null or fk2_uuid=null).
In this case, I can either define only entity_uuid as the partition key or entity_uuid, fk1_uuid, fk2_uuid combination as the partition key.
And this is a LOOKUP-type of table, meaning we don't plan to do any aggregations/slice-dice based on this table. And the rows will be rotated out since we will be inserting with TTL defined for each row.
Can someone enlighten me:
What is the downside of having too many partition keys with very few
rows in each? Is there a hit/cost on the storage engine level?
My understanding is the cluster keys are ALWAYS sorted. Does that mean having text columns in a cluster will always incur tree
balancing cost?
Well you can tell where my heart lies by now. However, when all rows in a partition all TTL-ed out, that partition still lives, or is there a way they will be removed by the DB engine as well?
Thanks,
Bing
The major and possibly most significant difference between having big partitions and small partitions is the ability to do range scans. If you want to be able to do scan queries like
SELECT * FROM table_1 where entity_id = x and fk1_uuid > something
Then you'll need to have the clustering column for performance, otherwise this query would be difficult (a multi-get at best, full table scan at worst.) I've never heard of any cases where having too many partitions is a drag on performance but having too wide a partition (ie lots of clustering column values) can cause issues when you get into the 1B+ cell range.
In terms of the cost of clustering, it is basically free at write time (in memory sort is very very fast) but you can incur costs at read time as partitions become spread amongst various SSTables. Small partitions which are written once will not occur the merge penalty since they will most likely only exist in 1 SSTable.
TTL'd partitions will be removed but be sure to read up on GC_GRACE_SECONDS to see how Cassandra actually deals with removing data.
TL;DR
Everything is dependent on your read/write pattern
No Range Scans? No need for clustering keys
Yes Range Scans? Clustering keys a must