voonhous opened a new issue, #20625:
URL: https://github.com/apache/hudi/issues/20625

   Snapshot (`_rt`) queries on MOR tables through the Trino connector are much 
slower than the same queries on COW or on the read-optimized view. Any file 
slice with log files leaves the native Parquet page source for the 
`HoodieFileGroupReader` merge path in 
`HudiPageSourceProvider.createPageSource`, and that path pays the costs below 
on every split.
   
   This issue tracks the work. Line references are against master `db90a0da`.
   
   ### Measured locally
   
   Generated TPC-DS-shaped MOR table: 48 partitions, 40k rows each, table 
version 9, 4 Avro log files per file group, 48 splits, storage behind injected 
object-store latency. Full scan with an aggregate:
   
   | | `_ro` | `_rt` |
   |---|---|---|
   | CPU | 0.1 - 0.3 s | about 9 s |
   | wall | 0.9 s | 3.8 - 4.0 s |
   | `hoodie.properties` reads | 2 | 50 (one per split) |
   | timeline LIST + instant GET | 1 | 97 (one of each per split) |
   
   Merge buffer alone, 300k log records of one file group: the size estimate is 
13,304 B/record, so 235,852 of 300,000 records (78.6%) spill to disk under the 
default 1 GiB budget.
   
   ### Work items
   
   - [ ] **Benchmarks.** Land the local MOR read benchmarks in `hudi-trino` 
test scope (`HudiMorReadBenchmark`, `HudiMergeBufferBenchmark`), so each item 
below has a number to move without a cluster.
   - [ ] **Record size estimator and spill serializer.** 
`HudiTrinoReaderContext` inherits `HoodieRecordSizeEstimator` and the Kryo 
`DefaultSerializer`, so every estimate includes the Avro schema object graph 
and the map spills early. `HoodieAvroReaderContext` already overrides both 
(`AvroRecordSizeEstimator`, `BufferedRecordSerializer`).
   - [ ] **Per-split meta client.** A new `HoodieTableMetaClient` is built per 
split (`HudiPageSourceProvider.java:232-234`, has a TODO), reading 
`hoodie.properties` each time, plus the index definitions file when one is 
registered. The coordinator already holds the table config; it could ship in 
the table handle (`TableState` is serializable and the file group reader 
accepts it).
   - [ ] **Per-split schema resolution.** With 
`hudi.table.resolve-column-name-casing.enabled=false` (the default) each split 
runs `TableSchemaResolver.getTableSchema()`: a recursive timeline listing, at 
least one instant file read with full commit metadata parsing, and an 
operation-field probe that can open a log file. Resolve once on the coordinator 
and ship it in the handle regardless of that flag.
   - [ ] **Log file sizes dropped.** `HudiUtil.convertToFileSlice` rebuilds 
each log file without the size the split already carries, so the reader stats 
every log file again; native log files also get two footer reads each.
   - [ ] **One page per file group.** `HudiPageSource.getNextSourcePage` drains 
the whole merged file group into a single page (no `pageBuilder.isFull()` 
check): no pipelining, a large transient heap, and `LIMIT` cannot stop early.
   - [ ] **Memory and I/O accounting.** `getMemoryUsage`, `getCompletedBytes` 
and `getReadTimeNanos` only cover the base file page source; the merge map and 
all log reads are invisible to Trino.
   - [ ] **Record-key predicate pushdown.** The base file is read with 
`TupleDomain.all()` and no dynamic filter when logs are present, so there is no 
row-group or page pruning. Spark and Flink push record-key-only filters to the 
base file under MOR; Trino pushes nothing. (Partition-level dynamic filtering 
in `HudiSplitSource` already applies to MOR splits.)
   - [ ] **Row-by-row Avro conversion.** Every base and log row becomes a 
`GenericData.Record` with boxed and decoded values, then is re-encoded into a 
page. Larger change; scope after the items above are measured.
   - [ ] **One split per file slice.** Slices with logs are never cut by size, 
which limits scan parallelism on large base files. Needs the file group reader 
to support `start != 0` with logs.
   
   Related: #20522, #19344.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to