What is the problem the feature request solves?
Every native shuffle block is a self-contained Arrow IPC stream: schema message, dictionary messages, record batch message, end-of-stream marker. On the read side read_single_batch builds a fresh StreamReader for each block, so the schema flatbuffer is parsed once per block, even though every block in a shuffle carries the same schema and the reducer already knows that schema from the plan protobuf (ShuffleScanExec builds it from data_types).
The write side already avoids the mirror image of this cost. ShuffleBlockWriter encodes the schema flatbuffer once in try_new and writes those pre-encoded bytes verbatim at the start of every block (SchemaEncoding::Precoded) rather than re-serializing per block. The reader has no equivalent.
Blocks per reduce task scale with the number of mappers times the number of blocks each mapper wrote for that partition, and a mapper writes at least one block per spill round, so this grows with both cluster width and memory pressure.
Describe the potential solution
The cost to remove is the flatbuffer parse and StreamReader setup, not I/O: the schema message sits inside the compressed stream and is decompressed either way. A reader that caches the parsed schema and skips re-parsing when a block's decompressed schema message matches the cached bytes would keep the current block format intact.
Dropping the schema message from the format outright would be cheaper still, since the reducer does not need it, but that changes the wire format and would have to account for the RSS path and for the dictionary-schema fallback, which does not use Precoded.
Unmeasured so far. Worth a read-side benchmark at a high mapper count before acting.
Additional context
No response
What is the problem the feature request solves?
Every native shuffle block is a self-contained Arrow IPC stream: schema message, dictionary messages, record batch message, end-of-stream marker. On the read side
read_single_batchbuilds a freshStreamReaderfor each block, so the schema flatbuffer is parsed once per block, even though every block in a shuffle carries the same schema and the reducer already knows that schema from the plan protobuf (ShuffleScanExecbuilds it fromdata_types).The write side already avoids the mirror image of this cost.
ShuffleBlockWriterencodes the schema flatbuffer once intry_newand writes those pre-encoded bytes verbatim at the start of every block (SchemaEncoding::Precoded) rather than re-serializing per block. The reader has no equivalent.Blocks per reduce task scale with the number of mappers times the number of blocks each mapper wrote for that partition, and a mapper writes at least one block per spill round, so this grows with both cluster width and memory pressure.
Describe the potential solution
The cost to remove is the flatbuffer parse and
StreamReadersetup, not I/O: the schema message sits inside the compressed stream and is decompressed either way. A reader that caches the parsed schema and skips re-parsing when a block's decompressed schema message matches the cached bytes would keep the current block format intact.Dropping the schema message from the format outright would be cheaper still, since the reducer does not need it, but that changes the wire format and would have to account for the RSS path and for the dictionary-schema fallback, which does not use
Precoded.Unmeasured so far. Worth a read-side benchmark at a high mapper count before acting.
Additional context
No response