-
Notifications
You must be signed in to change notification settings - Fork 61
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Add support for parallel reader and writer for ShardReplicationTask
- Loading branch information
Showing
10 changed files
with
222 additions
and
84 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
109 changes: 109 additions & 0 deletions
109
.../kotlin/com/amazon/elasticsearch/replication/task/shard/ShardReplicationChangesTracker.kt
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,109 @@ | ||
| package com.amazon.elasticsearch.replication.task.shard | ||
|
|
||
| import com.amazon.elasticsearch.replication.ReplicationSettings | ||
| import kotlinx.coroutines.delay | ||
| import kotlinx.coroutines.sync.Mutex | ||
| import kotlinx.coroutines.sync.withLock | ||
| import org.elasticsearch.common.logging.Loggers | ||
| import org.elasticsearch.index.shard.IndexShard | ||
| import java.util.Collections | ||
| import java.util.concurrent.atomic.AtomicLong | ||
| import kotlin.collections.ArrayList | ||
|
|
||
| /** | ||
| * Since we have added support for fetching batch of operations in parallel, we need to keep track of | ||
| * how many operations have been fetched and what batch needs to be fetched next. This creates the | ||
| * problem of concurrency with shared mutable state (https://kotlinlang.org/docs/shared-mutable-state-and-concurrency.html). | ||
| * ShardReplicationChangesTracker abstracts away all that complexity from ShardReplicationTask. | ||
| * Every reader coroutine in a shard has to interact with the tracker for: | ||
| * 1. Requesting the range of operations to be fetched in the batch. | ||
| * 2. Updating the final status of the batch fetch. | ||
| */ | ||
| class ShardReplicationChangesTracker(indexShard: IndexShard, private val replicationSettings: ReplicationSettings) { | ||
| private val log = Loggers.getLogger(javaClass, indexShard.shardId())!! | ||
|
|
||
| private val mutex = Mutex() | ||
| private val missingBatches = Collections.synchronizedList(ArrayList<Pair<Long, Long>>()) | ||
| private val observedSeqNoAtLeader = AtomicLong(indexShard.localCheckpoint) | ||
| private val seqNoAlreadyRequested = AtomicLong(indexShard.localCheckpoint) | ||
| private val batchSize = replicationSettings.batchSize | ||
|
|
||
| /** | ||
| * Provides a range of operations to be fetched next. | ||
| * | ||
| * Here are the guarantees that this method provides: | ||
| * 1. All reader coroutines get unique range of operations to fetch. | ||
| * 2. It'll ensure that the complete range of operations would be fetched. | ||
| * 3. Mutex in this method ensures that only one coroutine is requesting the batch at a time. | ||
| * If there are multiple coroutines, they'll be waiting in order to get the range of operations to fetch. | ||
| * 4. If we've already fetched all the operations from leader, there would be one and only one | ||
| * reader polling on leader per shard. | ||
| */ | ||
| suspend fun requestBatchToFetch():Pair<Long, Long> { | ||
| mutex.withLock { | ||
| logDebug("Waiting to get batch. requested: ${seqNoAlreadyRequested.get()}, leader: ${observedSeqNoAtLeader.get()}") | ||
|
|
||
| // Wait till we have batch to fetch. Note that if seqNoAlreadyRequested is equal to observedSeqNoAtLeader, | ||
| // we still should be sending one more request to fetch which will just do a poll and eventually timeout | ||
| // if no new operations are there on the leader (configured via TransportGetChangesAction.WAIT_FOR_NEW_OPS_TIMEOUT) | ||
| while (seqNoAlreadyRequested.get() > observedSeqNoAtLeader.get() && missingBatches.isEmpty()) { | ||
| delay(replicationSettings.pollDuration.millis) | ||
| } | ||
|
|
||
| // missing batch takes higher priority. | ||
| return if (missingBatches.isNotEmpty()) { | ||
| logDebug("Fetching missing batch ${missingBatches[0].first}-${missingBatches[0].second}") | ||
| missingBatches.removeAt(0) | ||
| } else { | ||
| // return the next batch to fetch and update seqNoAlreadyRequested. | ||
| val fromSeq = seqNoAlreadyRequested.getAndAdd(batchSize.toLong()) + 1 | ||
| val toSeq = fromSeq + batchSize - 1 | ||
| logDebug("Fetching the batch $fromSeq-$toSeq") | ||
| Pair(fromSeq, toSeq) | ||
| } | ||
| } | ||
| } | ||
|
|
||
| /** | ||
| * Ensures that we've successfully fetched a particular range of operations. | ||
| * In case of any failure(or we didn't get complete batch), we make sure that we're fetching the | ||
| * missing operations in the next batch. | ||
| */ | ||
| fun updateBatchFetched(success: Boolean, fromSeqNoRequested: Long, toSeqNoRequested: Long, toSeqNoReceived: Long, seqNoAtLeader: Long) { | ||
| if (success) { | ||
| // we shouldn't ever be getting more operations than requested. | ||
| assert(toSeqNoRequested >= toSeqNoReceived) { "${Thread.currentThread().getName()} Got more operations in the batch than requested" } | ||
| logDebug("Updating the batch fetched. ${fromSeqNoRequested}-${toSeqNoReceived}/${toSeqNoRequested}, seqNoAtLeader:$seqNoAtLeader") | ||
|
|
||
| // If we didn't get the complete batch that we had requested. | ||
| if (toSeqNoRequested > toSeqNoReceived) { | ||
| // If this is the last batch being fetched, update the seqNoAlreadyRequested. | ||
| if (seqNoAlreadyRequested.get() == toSeqNoRequested) { | ||
| seqNoAlreadyRequested.updateAndGet { toSeqNoReceived } | ||
| } else { | ||
| // Else, add to the missing operations to missing batch | ||
| logDebug("Didn't get the complete batch. Adding the missing operations ${toSeqNoReceived + 1}-${toSeqNoRequested}") | ||
| missingBatches.add(Pair(toSeqNoReceived + 1, toSeqNoRequested)) | ||
| } | ||
| } | ||
|
|
||
| // Update the sequence number observed at leader. | ||
| observedSeqNoAtLeader.getAndUpdate { value -> if (seqNoAtLeader > value) seqNoAtLeader else value } | ||
| logDebug("observedSeqNoAtLeader: ${observedSeqNoAtLeader.get()}") | ||
| } else { | ||
| // If this is the last batch being fetched, update the seqNoAlreadyRequested. | ||
| if (seqNoAlreadyRequested.get() == toSeqNoRequested) { | ||
| seqNoAlreadyRequested.updateAndGet { fromSeqNoRequested - 1 } | ||
| } else { | ||
| // If this was not the last batch, we might have already fetched other batch of | ||
| // operations after this. Adding this to missing. | ||
| logDebug("Adding batch to missing $fromSeqNoRequested-$toSeqNoRequested") | ||
| missingBatches.add(Pair(fromSeqNoRequested, toSeqNoRequested)) | ||
| } | ||
| } | ||
| } | ||
|
|
||
| private fun logDebug(msg: String) { | ||
| log.debug("${Thread.currentThread().name}: $msg") | ||
| } | ||
| } |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.