Github user squito commented on a diff in the pull request:
https://github.com/apache/spark/pull/15249#discussion_r81859662
--- Diff:
core/src/main/scala/org/apache/spark/scheduler/BlacklistTracker.scala ---
@@ -0,0 +1,130 @@
+/*
+ * 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.spark.scheduler
+
+import org.apache.spark.SparkConf
+import org.apache.spark.internal.config
+import org.apache.spark.internal.Logging
+import org.apache.spark.util.Utils
+
+private[scheduler] object BlacklistTracker extends Logging {
+
+ private val DEFAULT_TIMEOUT = "1h"
+
+ /**
+ * Returns true if the blacklist is enabled, based on checking the
configuration in the following
+ * order:
+ * 1. Is it specifically enabled or disabled?
+ * 2. Is it enabled via the legacy timeout conf?
+ * 3. Use the default for the spark-master:
+ * - off for local mode
+ * - on for distributed modes (including local-cluster)
+ */
+ def isBlacklistEnabled(conf: SparkConf): Boolean = {
+ conf.get(config.BLACKLIST_ENABLED) match {
+ case Some(isEnabled) =>
+ isEnabled
+ case None =>
+ // if they've got a non-zero setting for the legacy conf, always
enable the blacklist,
+ // otherwise, use the default based on the cluster-mode (off for
local-mode, on otherwise).
+ val legacyKey = config.BLACKLIST_LEGACY_TIMEOUT_CONF.key
+ conf.get(config.BLACKLIST_LEGACY_TIMEOUT_CONF) match {
+ case Some(legacyTimeout) =>
+ if (legacyTimeout == 0) {
+ logWarning(s"Turning off blacklisting due to legacy
configuaration:" +
+ s" $legacyKey == 0")
+ false
+ } else {
+ // mostly this is necessary just for tests, since real users
that want the blacklist
+ // will get it anyway by default
+ logWarning(s"Turning on blacklisting due to legacy
configuration:" +
+ s" $legacyKey > 0")
+ true
+ }
+ case None =>
+ // local-cluster is *not* considered local for these purposes,
we still want the
+ // blacklist enabled by default
+ !Utils.isLocalMaster(conf)
+ }
+ }
+ }
+
+ def getBlacklistTimeout(conf: SparkConf): Long = {
+ conf.get(config.BLACKLIST_TIMEOUT_CONF).getOrElse {
+ conf.get(config.BLACKLIST_LEGACY_TIMEOUT_CONF).getOrElse {
+ Utils.timeStringAsMs(DEFAULT_TIMEOUT)
+ }
+ }
+ }
+
+ /**
+ * Verify that blacklist configurations are consistent; if not, throw an
exception. Should only
+ * be called if blacklisting is enabled.
+ *
+ * The configuration for the blacklist is expected to adhere to a few
invariants. Default
+ * values follow these rules of course, but users may unwittingly change
one configuration
+ * without making the corresponding adjustment elsewhere. This ensures
we fail-fast when
+ * there are such misconfigurations.
+ */
+ def validateBlacklistConfs(conf: SparkConf): Unit = {
+
+ def mustBePos(k: String, v: String): Unit = {
+ throw new IllegalArgumentException(s"$k was $v, but must be > 0.")
+ }
+
+ // undocumented escape hatch for validation -- just for tests that
want to run in an "unsafe"
+ // configuration.
+ if (!conf.get("spark.blacklist.testing.skipValidation",
"false").toBoolean) {
+
+ Seq(
+ config.MAX_TASK_ATTEMPTS_PER_EXECUTOR,
+ config.MAX_TASK_ATTEMPTS_PER_NODE,
+ config.MAX_FAILURES_PER_EXEC_STAGE,
+ config.MAX_FAILED_EXEC_PER_NODE_STAGE
+ ).foreach { config =>
+ val v = conf.get(config)
+ if (v <= 0) {
+ mustBePos(config.key, v.toString)
+ }
+ }
+
+ val timeout = getBlacklistTimeout(conf)
+ if (timeout <= 0) {
+ // first, figure out where the timeout came from, to include the
right conf in the message.
+ conf.get(config.BLACKLIST_TIMEOUT_CONF) match {
+ case Some(t) =>
+ mustBePos(config.BLACKLIST_TIMEOUT_CONF.key, timeout.toString)
+ case None =>
+ mustBePos(config.BLACKLIST_LEGACY_TIMEOUT_CONF.key,
timeout.toString)
+ }
+ }
+
+ val maxTaskFailures = conf.getInt("spark.task.maxFailures", 4)
+ val maxNodeAttempts = conf.get(config.MAX_TASK_ATTEMPTS_PER_NODE)
+
+ if (maxNodeAttempts >= maxTaskFailures) {
--- End diff --
hmm I find your version a little misleading -- the problem is that
blacklist will fail to do its job, you'll abort the stage before you schedule
the task on a healthy location. Did you mean "... because blacklist will
_fail_ to kick in ..."?? But even that is doesn't seem right, because you also
say the task will *always* hit the maximum number of failures. It *might* work
in the unsafe configuration, but we can't make any guarantees.
Is it confusing because it doesn't clearly explain the mechanics of what
the end user should do differently? Or is it confusing because it doesn't
explain *why* this configuration is bad? I'm guessing the latter -- I hope
that at least what the user needs to do are clear.
I'm worried explaining why won't be very clear or will just be a huge blob
of text (or perhaps both). Perhaps the configuration docs need to be expanded
to explain this? Again I worry about a giant blob of text. But I do want to
try make the usage clear so definitely open to suggestions (if nothing else,
just selfishly, as I need to explain this to a lot of users).
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
