aaa8a80c9d
### What changes were proposed in this pull request?
Cache commonly occurring duplicate Some objects in SQLMetrics by using a Guava cache and reusing the existing Guava String Interner to avoid duplicate strings in JSONProtocol. Also with AccumulatorV2 we have seen lot of Some(-1L) and Some(0L) occurrences in a heap dump that is naively interned by having reusing a already constructed Some(-1L) and Some(0L)
To give some context on the impact and the garbage got accumulated, below are the details of the complex spark job which we troubleshooted and figured out the bottlenecks. **tl;dr - In short, major issues were the accumulation of duplicate objects mainly from SQLMetrics.**
Greater than 25% of the 40G driver heap filled with (a very large number of) **duplicate**, immutable objects.
1. Very large number of **duplicate** immutable objects.
- Type of metric is represented by `'scala.Some("sql")'` - which is created for each metric.
- Fixing this reduced memory usage from 4GB to a few bytes.
2. `scala.Some(0)` and `scala.Some(-1)` are very common metric values (typically to indicate absence of metric)
- Individually the values are all immutable, but spark sql was creating a new instance each time.
- Intern'ing these resulted in saving ~4.5GB for a 40G heap.
3. Using string interpolation for metric names.
- Interpolation results in creation of a new string object.
- We end up with a very large number of metric names - though the number of unique strings is miniscule.
- ~7.5 GB in the 40 GB heap : which went down to a few KB's when fixed.
### Why are the changes needed?
To reduce overall driver memory footprint which eventually reduces the Full GC pauses.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Since these are memory related optimizations, unit tests are not added. These changes are added in our internal platform which made it possible for one of the complex spark job continuously failing to succeed along with other set of optimizations.
Closes #32754 from venkata91/SPARK-35613.
Authored-by: Venkata krishnan Sowrirajan <vsowrirajan@linkedin.com>
Signed-off-by: Mridul Muralidharan <mridul<at>gmail.com>
|
||
|---|---|---|
| .. | ||
| catalyst | ||
| core | ||
| hive | ||
| hive-thriftserver | ||
| create-docs.sh | ||
| gen-sql-api-docs.py | ||
| gen-sql-config-docs.py | ||
| gen-sql-functions-docs.py | ||
| mkdocs.yml | ||
| README.md | ||
Spark SQL
This module provides support for executing relational queries expressed in either SQL or the DataFrame/Dataset API.
Spark SQL is broken up into four subprojects:
- Catalyst (sql/catalyst) - An implementation-agnostic framework for manipulating trees of relational operators and expressions.
- Execution (sql/core) - A query planner / execution engine for translating Catalyst's logical query plans into Spark RDDs. This component also includes a new public interface, SQLContext, that allows users to execute SQL or LINQ statements against existing RDDs and Parquet files.
- Hive Support (sql/hive) - Includes extensions that allow users to write queries using a subset of HiveQL and access data from a Hive Metastore using Hive SerDes. There are also wrappers that allow users to run queries that include Hive UDFs, UDAFs, and UDTFs.
- HiveServer and CLI support (sql/hive-thriftserver) - Includes support for the SQL CLI (bin/spark-sql) and a HiveServer2 (for JDBC/ODBC) compatible server.
Running ./sql/create-docs.sh generates SQL documentation for built-in functions under sql/site, and SQL configuration documentation that gets included as part of configuration.md in the main docs directory.