提交 e5b8546d 编写于 作者: G Gian Merlino

Autoscaling fixes.

- Initial targetWorkerCount must be subject to pool size limits
- Use consistent workerSetupData for the entire autoscaling run
- Don't call terminate() when we have nothing to terminate
- Terminate obsolete workers even faster
上级 c503f5e9
......@@ -104,7 +104,7 @@ public class SimpleResourceManagementStrategy implements ResourceManagementStrat
);
currentlyProvisioning.removeAll(workerNodeIds);
updateTargetWorkerCount(pendingTasks, zkWorkers);
updateTargetWorkerCount(workerSetupData, pendingTasks, zkWorkers);
int want = targetWorkerCount - (currValidWorkers + currentlyProvisioning.size());
while (want > 0) {
......@@ -180,16 +180,16 @@ public class SimpleResourceManagementStrategy implements ResourceManagementStrat
currentlyTerminating.clear();
currentlyTerminating.addAll(stillExisting);
updateTargetWorkerCount(pendingTasks, zkWorkers);
updateTargetWorkerCount(workerSetupData, pendingTasks, zkWorkers);
final Predicate<ZkWorker> isLazyWorker = createLazyWorkerPredicate(config, workerSetupData);
if (currentlyTerminating.isEmpty()) {
final int want = zkWorkers.size() - targetWorkerCount;
if (want > 0) {
final int excessWorkers = (zkWorkers.size() + currentlyProvisioning.size()) - targetWorkerCount;
if (excessWorkers > 0) {
final List<String> laziestWorkerIps =
FluentIterable.from(zkWorkers)
.filter(isLazyWorker)
.limit(want)
.limit(excessWorkers)
.transform(
new Function<ZkWorker, String>()
{
......@@ -202,19 +202,23 @@ public class SimpleResourceManagementStrategy implements ResourceManagementStrat
)
.toList();
log.info(
"Terminating %,d workers (wanted %,d): %s",
laziestWorkerIps.size(),
want,
Joiner.on(", ").join(laziestWorkerIps)
);
final AutoScalingData terminated = autoScalingStrategy.terminate(laziestWorkerIps);
if (terminated != null) {
currentlyTerminating.addAll(terminated.getNodeIds());
lastTerminateTime = new DateTime();
scalingStats.addTerminateEvent(terminated);
didTerminate = true;
if (laziestWorkerIps.isEmpty()) {
log.info("Wanted to terminate %,d workers, but couldn't find any lazy ones!");
} else {
log.info(
"Terminating %,d workers (wanted %,d): %s",
laziestWorkerIps.size(),
excessWorkers,
Joiner.on(", ").join(laziestWorkerIps)
);
final AutoScalingData terminated = autoScalingStrategy.terminate(laziestWorkerIps);
if (terminated != null) {
currentlyTerminating.addAll(terminated.getNodeIds());
lastTerminateTime = new DateTime();
scalingStats.addTerminateEvent(terminated);
didTerminate = true;
}
}
}
} else {
......@@ -283,12 +287,12 @@ public class SimpleResourceManagementStrategy implements ResourceManagementStrat
}
private void updateTargetWorkerCount(
final WorkerSetupData workerSetupData,
final Collection<RemoteTaskRunnerWorkItem> pendingTasks,
final Collection<ZkWorker> zkWorkers
)
{
synchronized (lock) {
final WorkerSetupData workerSetupData = workerSetupDataRef.get();
final Collection<ZkWorker> validWorkers = Collections2.filter(
zkWorkers,
createValidWorkerPredicate(config, workerSetupData)
......@@ -296,8 +300,14 @@ public class SimpleResourceManagementStrategy implements ResourceManagementStrat
final Predicate<ZkWorker> isLazyWorker = createLazyWorkerPredicate(config, workerSetupData);
if (targetWorkerCount < 0) {
// Initialize to size of current worker pool
targetWorkerCount = zkWorkers.size();
// Initialize to size of current worker pool, subject to pool size limits
targetWorkerCount = Math.max(
Math.min(
zkWorkers.size(),
workerSetupData.getMaxNumWorkers()
),
workerSetupData.getMinNumWorkers()
);
log.info(
"Starting with a target of %,d workers (current = %,d, min = %,d, max = %,d).",
targetWorkerCount,
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册