brijrajk commented on code in PR #12151:
URL: https://github.com/apache/gluten/pull/12151#discussion_r3600614771


##########
backends-velox/src/main/scala/org/apache/gluten/extension/RuntimeBloomFilterRewriteRule.scala:
##########
@@ -0,0 +1,78 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package org.apache.gluten.extension
+
+import org.apache.gluten.config.GlutenConfig
+import org.apache.gluten.expression.VeloxBloomFilterMightContain
+import org.apache.gluten.expression.aggregate.VeloxBloomFilterAggregate
+
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.catalyst.expressions.{BloomFilterMightContain, 
XxHash64}
+import 
org.apache.spark.sql.catalyst.expressions.aggregate.{AggregateExpression, 
BloomFilterAggregate}
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.execution.SparkPlan
+
+/**
+ * Physical pre-transform rule that rewrites runtime-filter bloom filters (the 
ones injected by
+ * Spark's `InjectRuntimeFilter` optimizer rule) to their Velox variants so 
they offload natively.
+ *
+ * Runtime bloom filters cannot be handled by 
[[BloomFilterMightContainJointRewriteRule]]: that rule
+ * is registered via `injectOptimizerRule`, which lands in Spark's Operator 
Optimization batch,
+ * while `InjectRuntimeFilter` runs in a later batch of `SparkOptimizer`. The 
runtime-filter
+ * expressions therefore do not exist yet when the logical rule fires, and a 
physical-level rewrite
+ * (as was always done before the logical rule was introduced) is required to 
keep
+ * `FilterExecTransformer` and the bloom-filter aggregate native.
+ *
+ * The rewrite is restricted to `InjectRuntimeFilter`'s exact expression 
shapes, which always wrap
+ * the key in [[XxHash64]] on both the producer and the consumer side:
+ *   - producer: `bloom_filter_agg(xxhash64(key), ...)` -> 
`velox_bloom_filter_agg(...)`
+ *   - consumer: `might_contain(bf, xxhash64(key))` -> 
`velox_might_contain(...)`
+ *
+ * Because each side is identifiable on its own, both are rewritten 
consistently to the Velox byte
+ * format (version=1) even when AQE compiles the bloom-filter subquery 
separately from the consuming
+ * filter stage. The `XxHash64` fingerprint also guarantees the other 
bloom-filter populations are
+ * never touched:
+ *   - `DataFrame.stat.bloomFilter()` builds `bloom_filter_agg(col, ...)` on 
the raw column (no
+ *     `XxHash64` wrapper) and deserializes the result with Spark's 
`BloomFilter.readFrom`, so its
+ *     bytes must stay in Spark-native format.
+ *   - User-facing `might_contain(<scalar subquery>, <value>)` pairs are 
already rewritten at the
+ *     logical level by [[BloomFilterMightContainJointRewriteRule]] (the 
GLUTEN-12013 fix), making
+ *     this rule a no-op for them.
+ *   - Literal-value pairs (SPARK-54336) contain no `XxHash64` and stay fully 
vanilla.
+ */
+case class RuntimeBloomFilterRewriteRule(spark: SparkSession) extends 
Rule[SparkPlan] {

Review Comment:
   @zhztheplayer @philo-he thanks for both proposals. Rather than answer from 
theory, I prototyped the `injectFinal` idea and probed it at runtime. The 
results changed my recommendation, so full findings below, including one 
correction to my own earlier comment.
   
   ## zhztheplayer's questions
   
   **"Why is the logical rule still kept?"** The Operator Optimization batch is 
a valid injection point for the expressions it targets: user-written 
`might_contain` exists at analysis time; only `InjectRuntimeFilter`'s 
expressions are missing there (created in a later `SparkOptimizer` batch). It 
must stay for two reasons:
   
   1. The literal-vs-non-literal pair decision (SPARK-54336 must stay vanilla; 
GLUTEN-12013 must become Velox) needs whole-plan visibility. After 
`PlanSubqueries` splits compilation, a producer-side `bloom_filter_agg(col)` 
alone is ambiguous between the two cases; there is no structural fingerprint 
like the runtime filters' `xxhash64` wrapper to disambiguate.
   2. Being logical bakes the rewrite into the plan before physical planning, 
so `originalPlan` itself carries the Velox variants; the user-facing population 
is reversion-safe with zero patching (proven by the threshold=1/2 tests).
   
   **"Any risk if we always replace in the physical phase / move to 
`injectFinal`?"** Yes, two, and the second is empirical:
   
   1. *Move* alone would regress offload: validation needs the Velox expression 
before `HeuristicTransform`, so it has to be keep-preTransform-plus-add-final.
   2. I implemented exactly that (same `xxhash64` guards, registered 
additionally at `injectFinal`) and reran the probes. Probe D flipped from crash 
to correct, and the guards held (`stat.bloomFilter`, SPARK-54336, 
`native.bloomFilter=false` all fine). But probe E 
(`wholeStage.fallback.threshold=1` alone) flipped from a crash to a **silent 
wrong result: 6 of 10 rows, deterministically missing the same keys across 
reruns**. A bloom pre-filter can only produce false positives, so missing rows 
means the filter itself was corrupt. That is strictly worse than the crash, so 
I have not pushed it.
   
   ## Root cause of the silent corruption (new finding)
   
   Plan diagnostics for probe E showed a **phase split**: the producer's 
partial aggregate stayed native 
(`FlushableHashAggregateExecTransformer[VeloxBloomFilterAggregate:Partial]`) 
while the reverted final-aggregation stage ran on the JVM 
(`ObjectHashAggregateExec[VeloxBloomFilterAggregate:Final]`). Same expression 
class, same byte format, no crash. But the two implementations size their 
blooms from different sources:
   
   - JVM `VeloxBloomFilterAggregate.createAggregationBuffer()` uses the 
expression's `estimatedNumItems` (for runtime filters, `InjectRuntimeFilter` 
sets it to the creation-side row count: 100 here).
   - Native `bloom_filter_agg` ignores the expression arguments and uses the 
session confs Gluten forwards to the native query context 
(`spark.sql.optimizer.runtime.bloomFilter.*`, defaults 1M items / 8388608 bits).
   
   The JVM final agg then merges a large native partial into its small buffer 
through Velox's `BloomFilter::merge` (common/base/BloomFilter.h), whose size 
guard is `VELOX_DCHECK_EQ(bits_.size(), otherSize)`, compiled out in release 
builds, followed by an unconditional `bits::orBits(..., 64 * otherSize)`. Bit 
positions were inserted modulo the large capacity and queried modulo the small 
one: deterministic false negatives.
   
   This incompatibility is latent and independent of the PoC: any execution 
pairing a native-partial with a JVM-final `velox_bloom_filter_agg` under 
divergent capacity parameters is exposed. Today the known reversion paths 
surface it as a crash because the vanilla/velox format check fails first; the 
PoC removed the crash and unmasked the silent path.
   
   **Correction to my Jul 14 comment:** probe E's original crash was this same 
phase split, i.e. the vanilla final aggregate choking on native partial 
buffers, not the consumer filter reverting (the consumer stayed a native 
`FilterExecTransformer` throughout). Same class of failure, wrong operator 
attributed.
   
   ## philo-he's proposals
   
   **Spark extension API for post-`InjectRuntimeFilter` logical rules.** Agreed 
this is the cleanest long-term shape: the runtime-filter rewrite would land in 
the optimized logical plan, so `originalPlan` carries it into every stage and 
reversion can never strip it, exactly like the user-facing path today. The 
physical rule would then be deleted entirely. Two qualifications: (a) as you 
noted, upstream acceptance is uncertain and it is a multi-release timeline, so 
it cannot be this PR's blocker; (b) the phase-split finding above means even 
this does not fully close the problem by itself. A reverted final-agg stage 
still executes the JVM implementation against native partial buffers, so the 
capacity alignment is a prerequisite regardless of where the rewrite runs. I 
think it is still worth filing the SPARK JIRA; happy to draft it.
   
   **Block whole-stage fallback when a stage contains a bloom filter.** The 
phase-split finding actually strengthens this option: it is the only proposal 
that prevents reversion-induced splits entirely (both the producer/consumer 
split and the partial/final split), sidestepping the format problem and the 
capacity problem at once. On implementation, the layering concern is solvable 
without `ExpandFallbackPolicy` learning Velox types: the policy is constructed 
in `VeloxRuleApi` (`injectFallbackPolicy(c => p => 
ExpandFallbackPolicy(...))`), so backends-velox can pass a bloom-detecting 
predicate into it. The trade-off is that it overrides the user's explicit 
fallback decision for those stages; whether that is acceptable is a judgement 
call I would defer to you both.
   
   ## Where this leaves the three options
   
   - **`injectFinal` (zhztheplayer):** right structure, but only safe after the 
capacity divergence is fixed. Prerequisite: make JVM 
`VeloxBloomFilterAggregate` size its buffer from the same source as native (the 
forwarded session confs), or adopt the incoming filter's size on first merge 
(`bits_.empty()` branch already supports adoption). Also worth proposing to 
Velox that the `DCHECK` become a user check so a mismatch can never be silent.
   - **Fallback-policy block (philo-he):** most robust short-term; no capacity 
prerequisite; costs an override of the user's fallback intent.
   - **Upstream extension point (philo-he):** best long-term; needs capacity 
alignment anyway; not a blocker for this PR.
   
   ## Proposed sequencing
   
   1. Merge this PR as-is: it fixes GLUTEN-12013 and SPARK-54336, restores 
native runtime filters, changes no goldens, and the residual reversion gap is 
default-off, fails loudly, and is byte-identical to main (verified by running 
the same probes against main's bloom code).
   2. Follow-up issue 1: capacity alignment between JVM and native 
`velox_bloom_filter_agg` (prerequisite), plus a phase-split regression test.
   3. Follow-up issue 2: close the reversion gap on top of (2), via 
`injectFinal` or the policy predicate, whichever you prefer.
   4. Long-term: SPARK JIRA for a post-`InjectRuntimeFilter` extension point, 
which would let the physical rule collapse into the logical one.
   
   Happy to file the follow-up issues with the probe reproductions. Does this 
sequencing work for you both?
   



-- 
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]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to