+ ShardDataTreeCohort createReadyCohort(final TransactionIdentifier txId, final DataTreeModification mod,
+ final java.util.Optional<SortedSet<String>> participatingShardNames) {
+ SimpleShardDataTreeCohort cohort = new SimpleShardDataTreeCohort(this, mod, txId,
+ cohortRegistry.createCohort(schemaContext, txId, shard::executeInSelf,
+ COMMIT_STEP_TIMEOUT), participatingShardNames);
+ pendingTransactions.add(new CommitEntry(cohort, readTime()));
+ return cohort;
+ }
+
+ // Exposed for ShardCommitCoordinator so it does not have deal with local histories (it does not care), this mimics
+ // the newReadWriteTransaction()
+ ShardDataTreeCohort newReadyCohort(final TransactionIdentifier txId, final DataTreeModification mod,
+ final java.util.Optional<SortedSet<String>> participatingShardNames) {
+ if (txId.getHistoryId().getHistoryId() == 0) {
+ return createReadyCohort(txId, mod, participatingShardNames);
+ }
+
+ return ensureTransactionChain(txId.getHistoryId(), null).createReadyCohort(txId, mod, participatingShardNames);
+ }
+
+ @SuppressFBWarnings(value = "DB_DUPLICATE_SWITCH_CLAUSES", justification = "See inline comments below.")
+ void checkForExpiredTransactions(final long transactionCommitTimeoutMillis,
+ final Function<SimpleShardDataTreeCohort, Optional<Long>> accessTimeUpdater) {
+ final long timeout = TimeUnit.MILLISECONDS.toNanos(transactionCommitTimeoutMillis);
+ final long now = readTime();
+
+ final Queue<CommitEntry> currentQueue = !pendingFinishCommits.isEmpty() ? pendingFinishCommits :
+ !pendingCommits.isEmpty() ? pendingCommits : pendingTransactions;
+ final CommitEntry currentTx = currentQueue.peek();
+ if (currentTx == null) {
+ // Empty queue, no-op
+ return;
+ }
+
+ long delta = now - currentTx.lastAccess;
+ if (delta < timeout) {
+ // Not expired yet, bail
+ return;
+ }
+
+ final Optional<Long> updateOpt = accessTimeUpdater.apply(currentTx.cohort);
+ if (updateOpt.isPresent()) {
+ final long newAccess = updateOpt.get().longValue();
+ final long newDelta = now - newAccess;
+ if (newDelta < delta) {
+ LOG.debug("{}: Updated current transaction {} access time", logContext,
+ currentTx.cohort.getIdentifier());
+ currentTx.lastAccess = newAccess;
+ delta = newDelta;
+ }
+
+ if (delta < timeout) {
+ // Not expired yet, bail
+ return;
+ }
+ }
+
+ final long deltaMillis = TimeUnit.NANOSECONDS.toMillis(delta);
+ final State state = currentTx.cohort.getState();
+
+ LOG.warn("{}: Current transaction {} has timed out after {} ms in state {}", logContext,
+ currentTx.cohort.getIdentifier(), deltaMillis, state);
+ boolean processNext = true;
+ final TimeoutException cohortFailure = new TimeoutException("Backend timeout in state " + state + " after "
+ + deltaMillis + "ms");
+
+ switch (state) {
+ case CAN_COMMIT_PENDING:
+ currentQueue.remove().cohort.failedCanCommit(cohortFailure);
+ break;
+ case CAN_COMMIT_COMPLETE:
+ // The suppression of the FindBugs "DB_DUPLICATE_SWITCH_CLAUSES" warning pertains to this clause
+ // whose code is duplicated with PRE_COMMIT_COMPLETE. The clauses aren't combined in case the code
+ // in PRE_COMMIT_COMPLETE is changed.
+ currentQueue.remove().cohort.reportFailure(cohortFailure);
+ break;
+ case PRE_COMMIT_PENDING:
+ currentQueue.remove().cohort.failedPreCommit(cohortFailure);
+ break;
+ case PRE_COMMIT_COMPLETE:
+ // FIXME: this is a legacy behavior problem. Three-phase commit protocol specifies that after we
+ // are ready we should commit the transaction, not abort it. Our current software stack does
+ // not allow us to do that consistently, because we persist at the time of commit, hence
+ // we can end up in a state where we have pre-committed a transaction, then a leader failover
+ // occurred ... the new leader does not see the pre-committed transaction and does not have
+ // a running timer. To fix this we really need two persistence events.
+ //
+ // The first one, done at pre-commit time will hold the transaction payload. When consensus
+ // is reached, we exit the pre-commit phase and start the pre-commit timer. Followers do not
+ // apply the state in this event.
+ //
+ // The second one, done at commit (or abort) time holds only the transaction identifier and
+ // signals to followers that the state should (or should not) be applied.
+ //
+ // In order to make the pre-commit timer working across failovers, though, we need
+ // a per-shard cluster-wide monotonic time, so a follower becoming the leader can accurately
+ // restart the timer.
+ currentQueue.remove().cohort.reportFailure(cohortFailure);
+ break;
+ case COMMIT_PENDING:
+ LOG.warn("{}: Transaction {} is still committing, cannot abort", logContext,
+ currentTx.cohort.getIdentifier());
+ currentTx.lastAccess = now;
+ processNext = false;
+ return;
+ case READY:
+ currentQueue.remove().cohort.reportFailure(cohortFailure);
+ break;
+ case ABORTED:
+ case COMMITTED:
+ case FAILED:
+ default:
+ currentQueue.remove();
+ }
+
+ if (processNext) {
+ processNextPending();
+ }
+ }
+
+ boolean startAbort(final SimpleShardDataTreeCohort cohort) {
+ final Iterator<CommitEntry> it = Iterables.concat(pendingFinishCommits, pendingCommits,
+ pendingTransactions).iterator();
+ if (!it.hasNext()) {
+ LOG.debug("{}: no open transaction while attempting to abort {}", logContext, cohort.getIdentifier());
+ return true;
+ }
+
+ // First entry is special, as it may already be committing
+ final CommitEntry first = it.next();
+ if (cohort.equals(first.cohort)) {
+ if (cohort.getState() != State.COMMIT_PENDING) {
+ LOG.debug("{}: aborting head of queue {} in state {}", logContext, cohort.getIdentifier(),
+ cohort.getIdentifier());
+
+ it.remove();
+ if (cohort.getCandidate() != null) {
+ rebaseTransactions(it, dataTree);
+ }
+
+ processNextPending();
+ return true;
+ }
+
+ LOG.warn("{}: transaction {} is committing, skipping abort", logContext, cohort.getIdentifier());
+ return false;
+ }
+
+ DataTreeTip newTip = MoreObjects.firstNonNull(first.cohort.getCandidate(), dataTree);
+ while (it.hasNext()) {
+ final CommitEntry e = it.next();
+ if (cohort.equals(e.cohort)) {
+ LOG.debug("{}: aborting queued transaction {}", logContext, cohort.getIdentifier());
+
+ it.remove();
+ if (cohort.getCandidate() != null) {
+ rebaseTransactions(it, newTip);
+ }
+
+ return true;
+ } else {
+ newTip = MoreObjects.firstNonNull(e.cohort.getCandidate(), newTip);
+ }
+ }
+
+ LOG.debug("{}: aborted transaction {} not found in the queue", logContext, cohort.getIdentifier());
+ return true;
+ }
+
+ @SuppressWarnings("checkstyle:IllegalCatch")
+ private void rebaseTransactions(final Iterator<CommitEntry> iter, final @NonNull DataTreeTip newTip) {
+ tip = Preconditions.checkNotNull(newTip);
+ while (iter.hasNext()) {
+ final SimpleShardDataTreeCohort cohort = iter.next().cohort;
+ if (cohort.getState() == State.CAN_COMMIT_COMPLETE) {
+ LOG.debug("{}: Revalidating queued transaction {}", logContext, cohort.getIdentifier());
+
+ try {
+ tip.validate(cohort.getDataTreeModification());
+ } catch (DataValidationFailedException | RuntimeException e) {
+ LOG.debug("{}: Failed to revalidate queued transaction {}", logContext, cohort.getIdentifier(), e);
+ cohort.reportFailure(e);
+ }
+ } else if (cohort.getState() == State.PRE_COMMIT_COMPLETE) {
+ LOG.debug("{}: Repreparing queued transaction {}", logContext, cohort.getIdentifier());
+
+ try {
+ tip.validate(cohort.getDataTreeModification());
+ DataTreeCandidateTip candidate = tip.prepare(cohort.getDataTreeModification());
+
+ cohort.setNewCandidate(candidate);
+ tip = candidate;
+ } catch (RuntimeException | DataValidationFailedException e) {
+ LOG.debug("{}: Failed to reprepare queued transaction {}", logContext, cohort.getIdentifier(), e);
+ cohort.reportFailure(e);
+ }
+ }
+ }
+ }
+
+ void setRunOnPendingTransactionsComplete(final Runnable operation) {
+ runOnPendingTransactionsComplete = operation;
+ maybeRunOperationOnPendingTransactionsComplete();
+ }
+
+ private void maybeRunOperationOnPendingTransactionsComplete() {
+ if (runOnPendingTransactionsComplete != null && !anyPendingTransactions()) {
+ LOG.debug("{}: Pending transactions complete - running operation {}", logContext,
+ runOnPendingTransactionsComplete);
+
+ runOnPendingTransactionsComplete.run();
+ runOnPendingTransactionsComplete = null;
+ }
+ }
+
+ ShardStats getStats() {
+ return shard.getShardMBean();
+ }
+
+ Iterator<SimpleShardDataTreeCohort> cohortIterator() {
+ return Iterables.transform(Iterables.concat(pendingFinishCommits, pendingCommits, pendingTransactions),
+ e -> e.cohort).iterator();
+ }
+
+ void removeTransactionChain(final LocalHistoryIdentifier id) {
+ if (transactionChains.remove(id) != null) {
+ LOG.debug("{}: Removed transaction chain {}", logContext, id);
+ }