hudi-agent commented on code in PR #19033:
URL: https://github.com/apache/hudi/pull/19033#discussion_r3463342744
##########
hudi-sync/hudi-hive-sync/src/main/java/org/apache/hudi/hive/HoodieHiveSyncClient.java:
##########
@@ -124,7 +137,10 @@ public HoodieHiveSyncClient(HiveSyncConfig config,
HoodieTableMetaClient metaCli
ddlExecutor = new HMSDDLExecutor(config, this.client);
break;
case HIVEQL:
- ddlExecutor = new HiveQueryDDLExecutor(config, this.client);
+ this.partitionDriverPool = maybeBuildHiveDriverPool(config);
Review Comment:
🤖 These two pools are built eagerly here, but if
`maybeBuildPartitionClientPool(...)` or the `HiveQueryDDLExecutor` constructor
on the next line throws, the outer `catch` (line ~148) just rethrows — and
since the constructor failed, `close()` is never called, so the already-built
pool's worker threads + Thrift connections leak. Could we close
`partitionDriverPool`/`partitionClientPool` in the catch before rethrowing?
Same pattern in the legacy default-mode branch below.
<sub><i>⚠️ AI-generated; verify before applying. React 👍/👎 to flag
quality.</i></sub>
##########
hudi-sync/hudi-hive-sync/src/main/java/org/apache/hudi/hive/HoodieHiveSyncClient.java:
##########
@@ -86,6 +90,15 @@ public class HoodieHiveSyncClient extends HoodieSyncClient {
private final Map<String, Table> initialTableByName = new HashMap<>();
DDLExecutor ddlExecutor;
private IMetaStoreClient client;
+ // Non-null only when HIVE_SYNC_BATCHING_ENABLED and sync mode is HIVEQL.
Owned by
+ // this class; closed in close() before Hive.closeCurrent().
HiveQueryDDLExecutor
+ // uses it only for DROP (Hive Thrift, not Hive Driver) — see
IMetaStoreClientPool
+ // javadoc.
+ private IMetaStoreClientPool partitionClientPool;
Review Comment:
🤖 nit: `partitionClientPool` is a raw nullable while `partitionDriverPool`
just below (line 101) is wrapped in `Option<>` — both represent the same
"batching enabled" condition, so it might be worth making the container type
consistent (either both `Option<>`, or both nullable with the same guard
pattern) to avoid a future reader wondering if there's a semantic difference.
<sub><i>⚠️ AI-generated; verify before applying. React 👍/👎 to flag
quality.</i></sub>
##########
hudi-sync/hudi-hive-sync/src/main/java/org/apache/hudi/hive/HiveSyncConfigHolder.java:
##########
@@ -122,6 +122,23 @@ public class HiveSyncConfigHolder {
.defaultValue(1000)
.markAdvanced()
.withDocumentation("The number of partitions one batch when synchronous
partitions to hive.");
+ public static final ConfigProperty<Boolean> HIVE_SYNC_BATCHING_ENABLED =
ConfigProperty
+ .key("hoodie.datasource.hive_sync.batching.enabled")
+ .defaultValue(false)
+ .markAdvanced()
+ .sinceVersion("1.1.0")
+ .withDocumentation("When true, partition operations
(add/update/touch/drop) in the active sync mode "
+ + "are split into batches of `hoodie.datasource.hive_sync.batch_num`
and dispatched in parallel "
+ + "to a pool of workers. Table-level operations (create/alter table,
last commit time, writer "
+ + "version) continue to use the single session client/driver.
Default off; existing behavior is "
+ + "unchanged unless explicitly opted in.");
+ public static final ConfigProperty<Integer> HIVE_SYNC_BATCHING_THREADS =
ConfigProperty
+ .key("hoodie.datasource.hive_sync.batching.threads")
+ .defaultValue(4)
+ .markAdvanced()
Review Comment:
🤖 nit: the doc says "Pool size (number of worker resources) and
worker-thread count" — could you simplify to something like "Number of worker
threads (and pooled metastore clients) for parallel partition sync when
`hoodie.datasource.hive_sync.batching.enabled` is true"? The current wording
lists "pool size" and "worker-thread count" as if they're distinct, but they're
the same number.
<sub><i>⚠️ AI-generated; verify before applying. React 👍/👎 to flag
quality.</i></sub>
##########
hudi-sync/hudi-hive-sync/src/main/java/org/apache/hudi/hive/util/HiveDriverPool.java:
##########
@@ -0,0 +1,340 @@
+/*
+ * 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.hudi.hive.util;
+
+import org.apache.hudi.exception.HoodieException;
+import org.apache.hudi.hive.HiveSyncConfig;
+import org.apache.hudi.hive.HoodieHiveSyncException;
+
+import org.apache.hadoop.hive.conf.HiveConf;
+import org.apache.hadoop.hive.ql.Driver;
+import org.apache.hadoop.hive.ql.session.SessionState;
+import org.apache.hadoop.security.UserGroupInformation;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.util.ArrayList;
+import java.util.List;
+import java.util.concurrent.CancellationException;
+import java.util.concurrent.ExecutionException;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+import java.util.concurrent.Future;
+import java.util.concurrent.ThreadFactory;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicInteger;
+
+import static
org.apache.hudi.sync.common.HoodieSyncConfig.META_SYNC_DATABASE_NAME;
+
+/**
+ * Pool of Hive {@link Driver} + {@link SessionState} pairs for parallel
HiveQL DDL.
+ *
+ * <p>Hive's {@code SessionState.start(state)} binds state to the calling
thread's
+ * thread-local, and {@code Driver} reads from that thread-local during {@code
run()}.
+ * A Driver constructed on one thread cannot be safely used from another. This
pool
+ * solves that by giving each slot its own dedicated worker thread (a
single-thread
+ * executor) — the Driver and SessionState are built on that thread by a
bootstrap
+ * task, and all subsequent SQL for that slot runs on the same thread.
+ *
+ * <p><b>Usage contract:</b> use this pool only for partition-row DDL
statements that
+ * are independent of each other and freely shuffleable across workers.
Table-level
+ * statements (createTable, schema evolution, USE database) must continue to
run on
+ * the session {@code Driver} held by {@code HiveQueryDDLExecutor} on the sync
driver
+ * thread. The pool is gated behind {@code
hoodie.datasource.hive_sync.batching.enabled}
+ * and is constructed only for HiveQL sync mode.
+ */
+public class HiveDriverPool implements AutoCloseable {
+
+ private static final Logger LOG =
LoggerFactory.getLogger(HiveDriverPool.class);
+
+ // Per-worker Driver construction has to be fast in practice (a few hundred
ms
+ // for the SessionState + Driver init). A 60s ceiling per worker leaves
plenty of
+ // headroom for a slow JVM warm-up but bounds the failure mode if the
metastore
+ // is unreachable or Hive hangs during init.
+ private static final long BOOTSTRAP_TIMEOUT_SECONDS = 60;
+
+ private final List<Worker> workers;
+ private final int size;
+ private volatile boolean closed;
+
+ public HiveDriverPool(HiveSyncConfig config, int size) {
+ this(config, size, new DefaultDriverFactory(config));
+ }
+
+ // Package-private for tests: accepts a DriverFactory so unit tests can
inject
+ // mock Driver instances without standing up a real Hive instance.
+ HiveDriverPool(HiveSyncConfig config, int size, DriverFactory factory) {
+ if (size < 1) {
+ throw new IllegalArgumentException("Pool size must be >= 1, got " +
size);
+ }
+ this.size = size;
+ this.workers = new ArrayList<>(size);
+ String databaseName = config.getStringOrDefault(META_SYNC_DATABASE_NAME);
+ PoolThreadFactory threadFactory = new PoolThreadFactory();
+ List<Future<Void>> bootstrapFutures = new ArrayList<>(size);
+ try {
+ for (int i = 0; i < size; i++) {
+ Worker worker = new Worker(threadFactory);
+ workers.add(worker);
+ bootstrapFutures.add(worker.executor.submit(() -> {
+ worker.driver = factory.newDriver(databaseName);
+ return null;
+ }));
+ }
+ // Block until all bootstraps complete so we surface construction errors
+ // before any caller hands us SQL. Bounded by BOOTSTRAP_TIMEOUT_SECONDS
so a
+ // hung Hive init doesn't deadlock the sync driver thread.
+ for (Future<Void> f : bootstrapFutures) {
+ f.get(BOOTSTRAP_TIMEOUT_SECONDS, TimeUnit.SECONDS);
+ }
+ } catch (Exception e) {
+ tearDown();
+ throw new HoodieException("Failed to construct HiveDriverPool of size "
+ size, e);
+ }
+ LOG.info("Initialized HiveDriverPool with {} workers", size);
+ }
+
+ /**
+ * Runs each given SQL on <i>every</i> worker, in order. Used for setup
statements
+ * (e.g. {@code USE database}) that must establish per-thread session context
+ * before any partition statement runs. Blocks until all workers have
completed
+ * the setup. Throws on first error.
+ */
+ public void runOnEachWorker(List<String> setupSqls) {
+ if (closed) {
+ throw new IllegalStateException("Cannot dispatch to a closed
HiveDriverPool");
+ }
+ if (setupSqls.isEmpty()) {
+ return;
+ }
+ List<Future<?>> futures = new ArrayList<>(workers.size());
+ for (Worker worker : workers) {
+ futures.add(worker.executor.submit(() -> {
+ for (String sql : setupSqls) {
+ worker.driver.run(sql);
+ }
+ return null;
+ }));
+ }
+ awaitAll(futures);
+ }
+
+ /**
+ * Dispatches each SQL string to a worker (round-robin) and returns the list
of
+ * futures. The caller is responsible for awaiting and collecting errors.
SQL text
+ * is intentionally not logged per-statement here: batched TOUCH/ADD
statements can
+ * be many kilobytes, and N parallel workers would multiply the log volume.
See
+ * {@link #awaitAll(List)} for the per-call summary log.
+ */
+ public List<Future<?>> runAll(List<String> sqls) {
+ if (closed) {
+ throw new IllegalStateException("Cannot dispatch to a closed
HiveDriverPool");
+ }
+ List<Future<?>> futures = new ArrayList<>(sqls.size());
+ for (int i = 0; i < sqls.size(); i++) {
+ String sql = sqls.get(i);
+ Worker worker = workers.get(i % workers.size());
+ futures.add(worker.executor.submit(() -> {
+ worker.driver.run(sql);
+ return null;
+ }));
+ }
+ return futures;
+ }
+
+ /**
+ * Awaits all futures and throws the first exception encountered. On first
failure,
+ * cancels the remaining (not yet started) futures so workers don't keep
running
+ * pointless work after a fatal error. Any errors that finished before
cancellation
+ * are logged at WARN. Callers do not need per-statement results (Hive's
Driver.run
+ * side-effects the metastore), so this method is void.
+ */
+ public void awaitAll(List<Future<?>> futures) {
+ long start = System.currentTimeMillis();
+ Exception firstError = null;
+ int completed = 0;
+ int cancelled = 0;
+ for (int i = 0; i < futures.size(); i++) {
+ Future<?> f = futures.get(i);
+ try {
+ f.get();
+ completed++;
+ } catch (CancellationException ce) {
+ // We cancelled this future ourselves after a prior error. Don't treat
it
+ // as a new failure; just note it for the summary log.
+ cancelled++;
+ } catch (InterruptedException ie) {
+ Thread.currentThread().interrupt();
+ if (firstError == null) {
+ firstError = ie;
+ cancelled += cancelRemaining(futures, i + 1);
+ }
+ } catch (ExecutionException ee) {
+ Exception cause = unwrap(ee);
+ if (firstError == null) {
+ firstError = cause;
+ cancelled += cancelRemaining(futures, i + 1);
+ } else {
+ LOG.warn("Additional SQL batch failed (suppressed in favor of first
error)", cause);
+ }
+ }
+ }
+ if (firstError != null) {
+ throw new HoodieHiveSyncException("Failed in executing SQL", firstError);
+ }
+ LOG.info("Completed {} SQL statements ({} cancelled) in {} ms across {}
workers",
+ completed, cancelled, System.currentTimeMillis() - start, size);
+ }
+
+ private static int cancelRemaining(List<Future<?>> futures, int fromIndex) {
+ int cancelled = 0;
+ for (int j = fromIndex; j < futures.size(); j++) {
+ // mayInterruptIfRunning=false: the worker thread is bound to a Hive
Driver
+ // whose state we don't want to corrupt mid-statement. Cancel only those
that
+ // haven't started yet; in-flight statements run to completion.
+ if (futures.get(j).cancel(false)) {
+ cancelled++;
+ }
+ }
+ return cancelled;
+ }
+
+ private static Exception unwrap(ExecutionException ee) {
+ Throwable cause = ee.getCause();
+ return (cause instanceof Exception) ? (Exception) cause : ee;
+ }
+
+ public int size() {
+ return size;
+ }
+
+ @Override
+ public void close() {
+ if (closed) {
+ return;
+ }
+ closed = true;
+ tearDown();
+ }
+
+ private void tearDown() {
+ // Close each worker's Driver/SessionState on its own thread, then shut the
+ // executor down. Running close() on the bound thread keeps SessionState's
+ // thread-local cleanup correct.
+ for (Worker worker : workers) {
+ try {
+ worker.executor.submit(() -> {
+ if (worker.driver != null) {
+ try {
+ worker.driver.close();
+ } catch (Exception e) {
+ LOG.warn("Error closing pooled Driver", e);
+ }
+ }
+ SessionState ss = SessionState.get();
+ if (ss != null) {
+ try {
+ ss.close();
+ } catch (Exception e) {
+ LOG.warn("Error closing pooled SessionState", e);
+ }
+ }
+ return null;
+ }).get(30, TimeUnit.SECONDS);
+ } catch (Exception e) {
+ LOG.warn("Error during pool worker shutdown", e);
+ }
+ worker.executor.shutdown();
+ try {
+ if (!worker.executor.awaitTermination(10, TimeUnit.SECONDS)) {
+ worker.executor.shutdownNow();
+ }
+ } catch (InterruptedException ie) {
+ worker.executor.shutdownNow();
+ Thread.currentThread().interrupt();
+ }
+ }
+ workers.clear();
+ }
+
+ /**
+ * Per-slot state: a single-thread executor and the Driver bound to its
thread.
+ * Driver is volatile because it is written by the bootstrap task and read by
+ * subsequent dispatch tasks on the same executor.
+ */
+ private static final class Worker {
+ final ExecutorService executor;
+ volatile Driver driver;
+
+ Worker(ThreadFactory threadFactory) {
+ this.executor = Executors.newSingleThreadExecutor(threadFactory);
+ }
+ }
+
+ @FunctionalInterface
+ interface DriverFactory {
+ Driver newDriver(String databaseName) throws Exception;
+ }
+
+ /**
+ * Builds a real Hive {@link Driver} on the calling thread. The SessionState
is
+ * constructed lazily (once, on the first worker thread that builds a
Driver) and
+ * shared across all worker threads — Hive uses ThreadLocal attachment, not
+ * exclusive ownership, so multiple workers calling
+ * {@code SessionState.start(sharedState)} all see the same config and
scratch dir
+ * without each spending the cost of building their own SessionState (and
risking
+ * resource-dir creation races during the constructor).
+ */
+ private static final class DefaultDriverFactory implements DriverFactory {
+ private final HiveConf hiveConf;
+ private volatile SessionState sharedSessionState;
+
+ DefaultDriverFactory(HiveSyncConfig config) {
+ this.hiveConf = config.getHiveConf();
+ }
+
+ @Override
+ public synchronized Driver newDriver(String databaseName) throws Exception
{
+ // SessionState is shared across workers; build it once on the first
call (with
+ // currentDatabase already set) and attach it to each worker's
thread-local on
+ // subsequent calls. The database is a pool-wide property and never
changes
+ // across workers, so setting it once at construction time is sufficient.
Review Comment:
🤖 Sharing one SessionState across worker threads is risky here. Hive stores
the active SessionState in a ThreadLocal (SessionState.start() sets
tss.set(...)), so this pattern installs the *same* mutable instance into N
threads' ThreadLocals. The Hive Driver isn't documented as thread-safe, and
Driver execution mutates session-scoped state — currentDatabase (USE), conf
overrides, and the compile/execute lineage/read-entity sets — so concurrent
ALTER PARTITION/TOUCH runs can interleave on those fields. Worth confirming
whether any of the batched statements can issue a USE or set conf; if so,
per-worker SessionState (or serializing the session-mutating bits) is the safer
default even at the init cost.
--
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]
