@Override
public synchronized DynamicPartitionMapImpl getMap() {
ClosableIterator<RowResult<Value>> iterator =
storage.getRange(PARTITION_MAP_TABLE, RangeRequest.all(), 1L);
Map<Cell, byte[]> cells = Maps.newHashMap();
while (iterator.hasNext()) {
RowResult<Value> row = iterator.next();
for (Entry<Cell, Value> entry : row.getCells()) {
assert !cells.containsKey(entry.getKey());
cells.put(entry.getKey(), entry.getValue().getContents());
}
}
iterator.close();
return DynamicPartitionMapImpl.fromTable(cells);
}
@Override
public SweepResults run(String tableName, int batchSize, @Nullable byte[] startRow) {
Preconditions.checkNotNull(tableName);
Preconditions.checkState(!AtlasDbConstants.hiddenTables.contains(tableName));
if (tableName.startsWith(AtlasDbConstants.NAMESPACE_PREFIX)) {
// this happens sometimes; I think it's because some places in the code can
// start this sweeper without doing the full normally ordered KVSModule startup.
// I did check and sweep.stats did contain the FQ table name for all of the tables,
// so it is at least broken in some way that still allows namespaced tables to eventually be
// swept.
log.warn("The sweeper should not be run on tables passed through namespace mapping.");
return SweepResults.EMPTY_SWEEP;
}
// Earliest start timestamp of any currently open transaction, with two caveats:
// (1) unreadableTimestamps are calculated via wall-clock time, and so may not be correct
// under pathological clock conditions
// (2) immutableTimestamps do not account for locks have timed out after checking their locks;
// such a transaction may have a start timestamp less than the immutableTimestamp, and it
// could still get successfully committed (its commit timestamp may or may not be less than
// the immutableTimestamp
// Note that this is fine, because we'll either
// (1) force old readers to abort (if they read a garbage collection sentinel), or
// (2) force old writers to retry (note that we must roll back any uncommitted transactions that
// we encounter
SweepStrategy sweepStrategy = sweepStrategyManager.get().get(tableName);
if (sweepStrategy == null) {
sweepStrategy = SweepStrategy.CONSERVATIVE;
} else if (sweepStrategy == SweepStrategy.NOTHING) {
return SweepResults.EMPTY_SWEEP;
}
if (startRow == null) {
startRow = new byte[0];
}
RangeRequest rangeRequest =
RangeRequest.builder().startRowInclusive(startRow).batchHint(batchSize).build();
long sweepTimestamp = getSweepTimestamp(sweepStrategy);
ClosableIterator<RowResult<Value>> valueResults;
if (sweepStrategy == SweepStrategy.CONSERVATIVE) {
valueResults = ClosableIterators.wrap(ImmutableList.<RowResult<Value>>of().iterator());
} else {
valueResults = keyValueService.getRange(tableName, rangeRequest, sweepTimestamp);
}
ClosableIterator<RowResult<Set<Long>>> rowResults =
keyValueService.getRangeOfTimestamps(tableName, rangeRequest, sweepTimestamp);
try {
List<RowResult<Set<Long>>> rowResultTimestamps =
ImmutableList.copyOf(Iterators.limit(rowResults, batchSize));
PeekingIterator<RowResult<Value>> peekingValues = Iterators.peekingIterator(valueResults);
Set<Cell> sentinelsToAdd = Sets.newHashSet();
Multimap<Cell, Long> rowTimestamps =
getTimestampsFromRowResults(rowResultTimestamps, sweepStrategy);
Multimap<Cell, Long> cellTsPairsToSweep =
getCellTsPairsToSweep(
rowTimestamps, peekingValues, sweepTimestamp, sweepStrategy, sentinelsToAdd);
sweepCells(tableName, cellTsPairsToSweep, sentinelsToAdd);
byte[] nextRow =
rowResultTimestamps.size() < batchSize
? null
: RangeRequests.getNextStartRow(
false, Iterables.getLast(rowResultTimestamps).getRowName());
return new SweepResults(nextRow, rowResultTimestamps.size(), cellTsPairsToSweep.size());
} finally {
rowResults.close();
valueResults.close();
}
}
