Spark Advanced Topics
Spark Advanced Topics Working Group Documentation
Details
SchemaColumnConvertNotSupportedException with vectorized Parquet reads
Bringing too much data back to the driver (collect and friends)
Too big broadcast joins
Tables getting broadcasted even when broadcast is disabled
Class or method not found
Container OOMs
spark.sql.AnalysisException: Correlated column is not allowed in predicate
Result size larger than spark.driver.maxResultSize error OR Kryo serialization failed: Buffer overflow.
Result size larger than spark.driver.maxResultsSize error
Driver ran out of memory
Driver ran out of memory
Executor out of disk error
Executor ran out of memory
Missing Files / File Not Found / Reading past RLE/BitPacking stream
Error
Memory Errors
Other errors
Fetch Failed exceptions
spark.sql.AnalysisException
Even Partitioning Yet Still Slow
Failed to read non-parquet file
Large record problems can show up in a few different ways.
Force computations
Key/Partition Skew
Max Serialized Task Size -- Serialized task A:B was X bytes, which exceeds max allowed: C spark.rpc.message.maxSize
Notenoughexecs
Partial v.s. Full Aggregates
PySpark UDF / UDAF OOM
Partition at read time
Bad Partitioning
Even Partitioning Yet Still Slow
ShuffleExchangeExec loses track of executor registrations
Corrupted Shuffle Blocks
Unexpected coalesce / shuffle down / SinglePartition, ENSURE_REQUIREMENTS
Slow executor
Slow job slow cluster
Slow job
Slow Map
Partition Filters
Slow reduce
Regular Expression Tips
Skewed Joins
Skewed/Slow Write
Identify the slow stage
Slow writes on S3
Slow writes due to Too many small files
Slow Writes
Too Big DAG (or when iterative algorithms go bump in the night)
Toofew tasks
Too Large JAR files
Toomany tasks
Avoid UDFs for the most part
Write Fails
Flowchart
Index
Error
Shared
Slow
Search
Previous
Next
Too many tasks
Search
From here you can search these documents. Enter your search terms below.
Keyboard Shortcuts
Keys
Action
?
Open this help
n
Next page
p
Previous page
s
Search