gortiz commented on code in PR #19407:
URL: https://github.com/apache/pinot/pull/19407#discussion_r3949957158


##########
pinot-broker/src/main/java/org/apache/pinot/broker/requesthandler/ServerPreConnector.java:
##########
@@ -0,0 +1,135 @@
+/**
+ * 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.pinot.broker.requesthandler;
+
+import com.google.common.annotations.VisibleForTesting;
+import com.google.common.util.concurrent.ThreadFactoryBuilder;
+import java.util.ArrayList;
+import java.util.Collection;
+import java.util.List;
+import java.util.concurrent.CompletionService;
+import java.util.concurrent.ExecutionException;
+import java.util.concurrent.ExecutorCompletionService;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+import java.util.concurrent.Future;
+import java.util.concurrent.TimeUnit;
+import java.util.function.BiPredicate;
+import java.util.function.Supplier;
+import javax.annotation.concurrent.ThreadSafe;
+import org.apache.pinot.core.transport.ServerInstance;
+import org.apache.pinot.spi.config.table.TableType;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+
+/// Opens broker-to-server Netty channels ahead of query traffic, so the first 
real query does not pay
+/// the blocking `connect()` -- and, when broker-to-server TLS is on, the 
handshake -- on its critical
+/// path.
+///
+/// `ServerRoutingInstance` identity includes the table type, so OFFLINE and 
REALTIME are **separate**
+/// channels to the same physical server; both are connected here. Connecting 
an already-active channel
+/// is a no-op, so this is safe to call more than once.
+///
+/// Bounded on two axes so it can never stall startup: a capped thread pool, 
and a per-channel wait
+/// clamped to the caller's deadline. A server that is unreachable or itself 
restarting is logged and
+/// skipped -- the existing lazy-connect path still serves it. This class is 
stateless and thread-safe.
+///
+/// It takes its dependencies as functions rather than concrete 
`RoutingManager`/`QueryRouter` types so
+/// the parallelism, budget and failure handling can be unit-tested without a 
live broker.
+@ThreadSafe
+public class ServerPreConnector {
+  private static final Logger LOGGER = 
LoggerFactory.getLogger(ServerPreConnector.class);
+
+  /// Cap on the connect thread pool: a large tenant must not spawn a thread 
per server. Safe to exceed
+  /// the core count even on a 2- or 4-vCPU broker: each task is blocking 
connect + TLS handshake (mostly
+  /// network wait, with the actual I/O on Netty's event loop), and this runs 
during startup before any
+  /// query load, so the threads are almost entirely parked rather than 
contending for CPU.
+  @VisibleForTesting
+  static final int MAX_CONNECT_THREADS = 16;
+
+  private final Supplier<Collection<ServerInstance>> _routableServersSupplier;
+  private final BiPredicate<ServerInstance, TableType> _connectFn;
+
+  /// @param routableServersSupplier supplies the servers to connect, 
evaluated once per [#preConnect]
+  ///     call after the caller has ensured routing is built
+  /// @param connectFn opens the channel for one (server, table type) and 
returns whether it succeeded
+  public ServerPreConnector(Supplier<Collection<ServerInstance>> 
routableServersSupplier,
+      BiPredicate<ServerInstance, TableType> connectFn) {
+    _routableServersSupplier = routableServersSupplier;
+    _connectFn = connectFn;
+  }
+
+  /// Opens a channel to every routable server, for both table types, in 
parallel, bounded by
+  /// `deadlineMs` (an absolute [System#currentTimeMillis] value). Returns the 
number of channels
+  /// successfully connected. Never throws: a channel that fails or times out 
is logged and skipped.
+  public int preConnect(long deadlineMs) {
+    // Snapshot the routable-server view once. The supplier may return a live 
map view that another thread
+    // updates during startup; snapshotting keeps the channel count consistent 
with the tasks actually
+    // submitted below, so we never poll for phantom channels or under-count 
real ones.
+    List<ServerInstance> servers = new 
ArrayList<>(_routableServersSupplier.get());
+    if (servers.isEmpty() || System.currentTimeMillis() >= deadlineMs) {
+      return 0;
+    }
+    long startMs = System.currentTimeMillis();
+    int channelCount = servers.size() * TableType.values().length;
+    ExecutorService executor = 
Executors.newFixedThreadPool(Math.min(channelCount, MAX_CONNECT_THREADS),

Review Comment:
   Your `connected == 0` variant is better than what I drafted, and the 
regression test names exactly why — an instantly-refused connect consuming 
iteration zero would have started the grace clock and abandoned the 
healthy-but-slower channels. Good catch.
   
   One follow-up, **non-blocking** — I'm approving regardless. Your tests cover 
many channels and slow channels separately, but never together, and that 
combination is where the grace window misreads the signal. I compiled 
`ServerPreConnector` from this head and ran both at once, with **every server 
healthy and nothing stuck**:
   
   | targets | per-connect | connected | elapsed |
   |---|---|---|---|
   | 16 | 3000 ms | 16/16 | 3024 ms |
   | **48** | **3000 ms** | **16/48** | **5011 ms** |
   | 48 | 100 ms | 48/48 | 316 ms |
   
   With more channels than workers the surplus completes in waves one 
connect-latency apart, so the quiet window the heuristic reads as *stuck* is 
really just the gap between waves. Two thirds of a healthy cluster drops out of 
the warm-up, and the WARN attributes it to a straggler grace window when 
nothing was straggling.
   
   The effect is benign — `shutdown()` lets the queued tasks finish and 
publish, so those channels still warm up in the background, and the broker is 
serving by then. What's lost is that the feature quietly under-delivers on any 
cluster whose connects are slower than the window, and the log misexplains it. 
Given `MAX_CONNECT_THREADS = 16`, a tenant of more than 16 (server, table type) 
pairs is ordinary, so this isn't only a pathological shape.
   
   Scaling the window off the first observed connect latency covers it. All 
tasks start together, so the first success approximates one connect's latency:
   
   ```diff
        int connected = 0;
        boolean releasedEarly = false;
   +    long graceMs = STRAGGLER_GRACE_MS;
        try {
   ...
   -        long waitMs = connected == 0 ? remainingMs : Math.min(remainingMs, 
STRAGGLER_GRACE_MS);
   +        long waitMs = connected == 0 ? remainingMs : Math.min(remainingMs, 
graceMs);
            try {
              Future<Boolean> future = completionService.poll(waitMs, 
TimeUnit.MILLISECONDS);
   ...
              if (Boolean.TRUE.equals(future.get())) {
   +            if (connected == 0) {
   +              // All tasks started together, so the first success 
approximates one connect's latency. With
   +              // more channels than workers the surplus completes in waves 
one latency apart, so a window
   +              // narrower than that latency would abandon healthy channels 
still queued behind the pool.
   +              graceMs = Math.max(STRAGGLER_GRACE_MS, 2 * 
(System.currentTimeMillis() - startMs));
   +            }
                connected++;
              }
   ```
   
   Verified against both shapes, so the straggler behaviour this thread was 
about is untouched:
   
   | scenario | as pushed | with the scaled window |
   |---|---|---|
   | 48 healthy @ 3000 ms | 16/48 in 5011 ms | **48/48 in 9013 ms** |
   | 48 healthy @ 100 ms | 48/48 in 316 ms | 48/48 in 315 ms |
   | 8 channels, one black-holed | 7/8 in 2022 ms | 7/8 in **2022 ms** |
   | 48 channels, one black-holed | 47/48 in 2008 ms | 47/48 in **2006 ms** |
   
   Entirely your call whether to take it now or leave it — the current 
behaviour degrades safely, so it isn't a merge blocker. If you do take it, a 
test with `MAX_CONNECT_THREADS * 3` targets each slower than 
`STRAGGLER_GRACE_MS` would pin it; that's the shape that produced the 16/48 
above.



-- 
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