je/test/standalone/FailoverTest.java

562 lines
19 KiB
Java
Raw Normal View History

2021-06-06 17:46:45 +00:00
/*-
* Copyright (C) 2002, 2017, Oracle and/or its affiliates. All rights reserved.
*
* This file was distributed by Oracle as part of a version of Oracle Berkeley
* DB Java Edition made available at:
*
* http://www.oracle.com/technetwork/database/database-technologies/berkeleydb/downloads/index.html
*
* Please see the LICENSE file included in the top-level directory of the
* appropriate version of Oracle Berkeley DB Java Edition for a copy of the
* license and additional information.
*/
import java.io.File;
import java.util.Random;
import java.util.concurrent.TimeUnit;
import java.util.logging.Logger;
import com.sleepycat.je.EnvironmentConfig;
import com.sleepycat.je.Transaction;
import com.sleepycat.je.rep.RepInternal;
import com.sleepycat.je.rep.ReplicatedEnvironment;
import com.sleepycat.je.rep.ReplicationConfig;
import com.sleepycat.je.rep.utilint.RepTestUtils;
import com.sleepycat.je.rep.utilint.RepTestUtils.RepEnvInfo;
import com.sleepycat.je.utilint.LoggerUtils;
import com.sleepycat.je.utilint.TracerFormatter;
import com.sleepycat.je.utilint.VLSN;
import com.sleepycat.persist.EntityStore;
import com.sleepycat.persist.PrimaryIndex;
/**
* Exercises replica-only, master-only, and replica/master hybrid failover.
*/
public class FailoverTest {
private static final boolean verbose = Boolean.getBoolean("verbose");
private static final TracerFormatter dateFormat = new TracerFormatter();
/* Time to wait for each node to syncup to the master. */
private static final long NODE_SYNC_MS = 60 * 1000;
private RepEnvInfo[] repEnvInfo;
/* -------------------Configurable params----------------*/
/* Environment home root for whole replication group. */
private File envRoot;
/* Replication group size. */
private int nNodes = 5;
/* Database size. */
private int dbSize = 2000;
/* Steady state will finish after doing this number of transactions. */
private int steadyTxns = 60000;
/*
* Specify the JE log file size and checkpoint bytes to provoke log
* cleaning.
*/
private long logFileSize = 409600;
private long checkpointBytes = 1000000;
/*
* More than enough disk space to hold the data set and allow cleaning.
* Note that repeated syncups can prevent file deletion and more space
* may be needed for outlier cases.
*/
private long maxDisk = 500 * 1000000;
/* Select a new master after doing this number of operations. */
private int txnsPerRound = 100;
private static final int opsPerTxn = 20;
/* Cycle through the array nodes when provoking failovers. */
private int roundRobinIdx;
/* A value to use to create new records. */
private int nextVal = dbSize + 1;
/* Used for generating random keys. */
private final Random random = new Random();
/* Determines whether master, replica, or hybrid failover is tested. */
private FailoverAgent failoverAgent;
private int subDir = 0;
private boolean offHeap = false;
private Logger logger =
LoggerUtils.getLoggerFixedPrefix(getClass(), "FailoverTest");
private FailoverTest() {
}
public static void main(String args[]) {
try {
FailoverTest test = new FailoverTest();
test.parseArgs(args);
test.doRampup();
test.doSteadyState();
if (verbose) {
System.err.println("Test done");
}
} catch (Throwable t) {
t.printStackTrace(System.err);
System.exit(1);
}
}
/**
* Grow the data store to the appropriate size for the steady state
* portion of the test.
*/
private void doRampup()
throws Exception {
/*
* Clean up from previous runs. This test will not succeed if there is
* anything in the test environments.
*/
RepTestUtils.removeRepEnvironments(envRoot);
final long mainCacheSize;
final long offHeapCacheSize;
if (offHeap) {
mainCacheSize = nNodes * 10 * 1024 * 1024;
offHeapCacheSize = 100 * 1024 * 1024;
} else {
mainCacheSize = 0;
offHeapCacheSize = 0;
}
EnvironmentConfig envConfig = Utils.createEnvConfig(
logFileSize, maxDisk, checkpointBytes, subDir,
mainCacheSize, offHeapCacheSize);
/*
* We have a lot of environments open in a single process, so reduce
* the cache size lest we run out of file descriptors.
*/
envConfig.setConfigParam(EnvironmentConfig.LOG_FILE_CACHE_SIZE, "30");
repEnvInfo = RepTestUtils.setupEnvInfos(envRoot, nNodes, envConfig);
/* Increase timeout to avoid InsufficientAcksException. */
for (RepEnvInfo info : repEnvInfo) {
info.getRepConfig().setReplicaAckTimeout(50, TimeUnit.SECONDS);
}
if (subDir > 0) {
RepTestUtils.createRepSubDirs(repEnvInfo, subDir);
}
ReplicatedEnvironment master = Utils.getMaster(repEnvInfo);
EntityStore store = Utils.openStore(master, Utils.DB_NAME);
RepTestData.insertData(store, dbSize);
Utils.doSyncAndCheck(repEnvInfo);
}
/**
* Execute a steady stream of work until the steady state phase is over.
*/
private void doSteadyState()
throws Exception {
if (verbose) {
System.err.println("Steady state starting");
}
int round = 0;
failoverAgent.init();
while (true) {
round++;
if (verbose) {
System.err.println ("Round " + round);
}
/*
* If doWork returns false, it means the steady state stage is
* over. In this loop, the master stays up. The replicas are
* killed in a round robin fashion. No need to sync up between
* rounds; avoiding the syncup and check of data makes for more
* variation.
*/
if (!oneRoundWorkAndFailover(round)) {
if (verbose) {
System.err.println("Steady state is over, ending test.");
}
break;
}
}
/*
* Shutdown all nodes and sync them up using DbSync. Then check
* for node equality.
*/
ReplicatedEnvironment master = Utils.getMaster(repEnvInfo);
VLSN lastTxnEnd = RepInternal.getNonNullRepImpl(master).getVLSNIndex().
getRange().getLastTxnEnd();
for (RepEnvInfo info : repEnvInfo) {
info.closeEnv();
}
/*
* Run DbSync for each node. This is done concurrently in a thread per
* node, but the syncups are competing for resources, so the time to
* syncup and gain consistency for one node could in the worst case be
* as long as the time to do it serially for all nodes.
*/
long maxTimeoutMs = NODE_SYNC_MS * nNodes;
String maxTimeoutStr = String.valueOf(maxTimeoutMs) + " ms";
for (RepEnvInfo info : repEnvInfo) {
info.getRepConfig().setConfigParam(
ReplicationConfig.ENV_CONSISTENCY_TIMEOUT, maxTimeoutStr);
info.getRepConfig().setConfigParam(
ReplicationConfig.ENV_SETUP_TIMEOUT, maxTimeoutStr);
}
RepTestUtils.syncupGroup(maxTimeoutMs, repEnvInfo);
/* Re-open them to check node equality. */
for (RepEnvInfo info : repEnvInfo) {
info.getEnvConfig().setConfigParam("je.env.runCleaner", "false");
info.getEnvConfig().setConfigParam("je.env.runINCompressor",
"false");
}
RepTestUtils.restartGroup(repEnvInfo);
RepTestUtils.checkNodeEquality(lastTxnEnd, verbose, repEnvInfo);
RepTestUtils.checkUtilizationProfile(repEnvInfo);
for (RepEnvInfo info : repEnvInfo) {
info.closeEnv();
}
/* Use DbSpace to check utilization. */
}
/**
* One round of work: the master executes updates. The replicas are killed
* off in the middle of a round, so that they die on a non-commit boundary,
* and need to do some rollback.
*
* @return false if the rampup stage has finished.
*/
private boolean oneRoundWorkAndFailover(int round)
throws Exception {
boolean runAble = true;
// Utils.getMaster re-opens the whole group. Maybe we should try
// letting the killed nodes stay down longer.
ReplicatedEnvironment master = Utils.getMaster(repEnvInfo);
for (RepEnvInfo rep : repEnvInfo) {
logger.info(
RepInternal.getNonNullRepImpl(rep.getEnv()).dumpState());
}
EntityStore dbStore = Utils.openStore(master, Utils.DB_NAME);
PrimaryIndex<Integer, RepTestData> primaryIndex =
dbStore.getPrimaryIndex(Integer.class, RepTestData.class);
for (int whichTxn = 0; whichTxn < txnsPerRound; whichTxn++) {
Transaction txn = master.beginTransaction(null, null);
for (int i = 0; i < opsPerTxn; i++) {
/* Do a random update here. */
int key = random.nextInt(dbSize);
RepTestData data = new RepTestData();
data.setKey(key);
data.setData(nextVal++);
data.setName(RepTestData.generateNameField(key));
primaryIndex.put(txn, data);
}
/*
* Kill some set of nodes once during this round, in the middle
* of a transaction.
*/
if (failoverAgent.killOnIteration(whichTxn)) {
roundRobinIdx = failoverAgent.shutdownNodes(round,
whichTxn,
roundRobinIdx);
}
if (master.isValid()) {
/*
* This txn may be invalid if the master was killed by the
* failover agent.
*/
txn.commit();
}
/* Check whether the steady stage should break. */
if (--steadyTxns == 0) {
runAble = false;
break;
}
}
try {
dbStore.close();
} catch (RuntimeException e) {
if (master.isValid()) {
throw e;
}
}
return runAble;
}
private void setFailoverAgent(String modeValue)
throws IllegalArgumentException {
if (modeValue.equalsIgnoreCase("replica")) {
failoverAgent = new ReplicaFailover();
} else if (modeValue.equalsIgnoreCase("master")) {
failoverAgent = new MasterFailover();
} else if (modeValue.equalsIgnoreCase("hybrid")) {
failoverAgent = new HybridFailover();
} else {
throw new IllegalArgumentException
(modeValue +
" is not a legal value for -mode. " +
" Only master | replica | hybrid is accepted.");
}
}
private void parseArgs(String args[])
throws Exception {
for (int i = 0; i < args.length; i++) {
boolean moreArgs = i < args.length - 1;
if (args[i].equals("-h") && moreArgs) {
envRoot = new File(args[++i]);
} else if (args[i].equals("-repGroupSize") && moreArgs) {
nNodes = Integer.parseInt(args[++i]);
} else if (args[i].equals("-dbSize") && moreArgs) {
dbSize = Integer.parseInt(args[++i]);
} else if (args[i].equals("-logFileSize") && moreArgs) {
logFileSize = Long.parseLong(args[++i]);
} else if (args[i].equals("-maxDisk") && moreArgs) {
maxDisk = Long.parseLong(args[++i]);
} else if (args[i].equals("-steadyTxns") && moreArgs) {
steadyTxns = Integer.parseInt(args[++i]);
} else if (args[i].equals("-txnsPerRound") && moreArgs) {
txnsPerRound = Integer.parseInt(args[++i]);
} else if (args[i].equals("-checkpointBytes") && moreArgs) {
checkpointBytes = Long.parseLong(args[++i]);
} else if (args[i].equals("-mode") && moreArgs) {
setFailoverAgent(args[++i]);
} else if (args[i].equals("-subDir") && moreArgs) {
subDir = Integer.parseInt(args[++i]);
} else if (args[i].equals("-offheap") && moreArgs) {
offHeap = Boolean.parseBoolean(args[++i]);
} else {
usage("Unknown arg: " + args[i]);
}
}
if (nNodes <= 2) {
throw new IllegalArgumentException
("Replication group size should > 2!");
}
if (steadyTxns < txnsPerRound) {
throw new IllegalArgumentException
("steadyTxns should be larger than txnsPerRound!");
}
if (failoverAgent == null) {
usage("-mode must be specified");
}
}
private void usage(String error) {
if (error != null) {
System.err.println(error);
}
System.err.println
("java " + getClass().getName() + "\n" +
" [-h <replication group Environment home dir>]\n" +
" [-mode <master | replica | hybrid>\n" +
" [-repGroupSize <replication group size>]\n" +
" [-dbSize <records' number of the tested database>]\n" +
" [-logFileSize <JE log file size>]\n" +
" [-checkpointBytes <Checkpointer wakeup interval bytes>]\n" +
" [-steadyTxns <the total update operations in steady state>]\n" +
" [-txnsPerRound <select a new master after running this " +
" number of txns>]\n" +
" [-subDir <num directories to use with je.log.nDataDirectories");
System.exit(2);
}
/**
* The FailoverAgent decides which nodes to kill, and when to kill
* them.
*/
private abstract class FailoverAgent {
int victimsPerRound;
public void init() {
/*
* victimsPerRound Must be set after all args are parsed, to ensure
* that nNodes has been initialized.
*/
victimsPerRound = nNodes - RepTestUtils.getQuorumSize(nNodes);
}
/**
* @return true if a kill should be done on this iteration of the
* test loop.
*/
public abstract boolean killOnIteration(int currentTxnIndex);
/**
* @return the next index to start a shutdown, so that the test runs
* in a round robin fashion.
*/
public abstract int shutdownNodes(int whichRound,
int currentTxnIndex,
int lastShutdownIndex);
/**
* Generate a handy label which says which round and txn we are
* on, for debugging.
*/
String killLabel(int whichRound, int currentTxnIndex) {
return "Round " + whichRound +
" countdownTxn " + steadyTxns +
" killTxn " + currentTxnIndex + ":";
}
void killMaster(String label) {
for (RepEnvInfo info : repEnvInfo) {
if (info.isMaster()) {
if (verbose) {
System.err.println(
dateFormat.getDate(System.currentTimeMillis()) +
" " + label + " killing master " +
info.getEnv().getNodeName());
}
info.abnormalCloseEnv();
/*
* Be sure to break out of the loop after killing a master,
* because an election will start, and a new master might
* be elected as we continue to iterate over the set.
*/
break;
}
}
}
int killReplicas(String label,
int numVictims,
int lastShutdownIndex) {
int nShutdown = numVictims;
int shutdownIndex = lastShutdownIndex;
while (nShutdown > 0) {
RepEnvInfo info = repEnvInfo[shutdownIndex];
if (info.isReplica()) {
if (verbose) {
System.err.println(
dateFormat.getDate(System.currentTimeMillis()) +
" " + label + " killing replica " +
info.getEnv().getNodeName());
}
info.abnormalCloseEnv();
nShutdown--;
}
shutdownIndex++;
if (shutdownIndex == nNodes) {
shutdownIndex = 0;
}
}
return shutdownIndex;
}
}
/*
* Shutdown the maximum number of replicas that still leaves a quorum.
*/
private class ReplicaFailover extends FailoverAgent {
@Override
public boolean killOnIteration(int currentTxn) {
/* Always kill in the middle of the first txn of the round. */
return currentTxn == 0;
}
@Override
public int shutdownNodes(int whichRound,
int currentTxnIndex,
int lastShutdownIndex) {
return killReplicas(killLabel(whichRound, currentTxnIndex),
victimsPerRound, lastShutdownIndex);
}
}
/*
* Shutdown the master only.
*/
private class MasterFailover extends FailoverAgent {
@Override
public boolean killOnIteration(int currentTxn) {
/*
* Always kill in the middle of the last txn of the round. Since
* this is the master, we need to have it alive to execute writes.
*/
return currentTxn == (txnsPerRound - 1);
}
@Override
public int shutdownNodes(int whichRound,
int currentTxnIndex,
int lastShutdownIndex) {
killMaster(killLabel(whichRound, currentTxnIndex));
return 0;
}
}
/*
* Shutdown the the master and in addition the maximum number of replicas
* that still leaves a quorum.
*/
private class HybridFailover extends FailoverAgent {
@Override
public boolean killOnIteration(int currentTxn) {
/*
* Always kill in the middle of the last txn of the round. Since
* this is the master, we need to have it alive to execute writes.
*/
return currentTxn == txnsPerRound - 1;
}
@Override
public int shutdownNodes(int whichRound,
int currentTxnIndex,
int lastShutdownIndex) {
String label = killLabel(whichRound, currentTxnIndex);
killMaster(label);
/*
* Kill one less than the prescribed victims per round, because
* we already killed the master.
*/
return killReplicas(label, victimsPerRound-1, lastShutdownIndex);
}
}
}