{
logger.info("Writing " + Memtable.this.toString());
SSTableReader ssTable;
// errors when creating the writer that may leave empty temp files.
SSTableWriter writer = createFlushWriter(cfs.getTempSSTablePath(sstableDirectory));
try
{
// (we can't clear out the map as-we-go to free up memory,
// since the memtable is being used for queries in the "pending flush" category)
for (Map.Entry<RowPosition, AtomicSortedColumns> entry : rows.entrySet())
{
ColumnFamily cf = entry.getValue();
if (cf.isMarkedForDelete())
{
// When every node is up, there's no reason to write batchlog data out to sstables
// (which in turn incurs cost like compaction) since the BL write + delete cancel each other out,
// and BL data is strictly local, so we don't need to preserve tombstones for repair.
// If we have a data row + row level tombstone, then writing it is effectively an expensive no-op so we skip it.
// See CASSANDRA-4667.
if (cfs.name.equals(SystemKeyspace.BATCHLOG_CF) && cfs.keyspace.getName().equals(Keyspace.SYSTEM_KS) && !(cf.getColumnCount() == 0))
continue;
// Pedantically, you could purge column level tombstones that are past GcGRace when writing to the SSTable.
// But it can result in unexpected behaviour where deletes never make it to disk,
// as they are lost and so cannot override existing column values. So we only remove deleted columns if there
// is a CF level tombstone to ensure the delete makes it into an SSTable.
ColumnFamilyStore.removeDeletedColumnsOnly(cf, Integer.MIN_VALUE);
}
writer.append((DecoratedKey)entry.getKey(), cf);
}
if (writer.getFilePointer() > 0)
{
ssTable = writer.closeAndOpenReader();
logger.info(String.format("Completed flushing %s (%d bytes) for commitlog position %s",
ssTable.getFilename(), new File(ssTable.getFilename()).length(), context.get()));
}
else
{
writer.abort();
ssTable = null;
logger.info("Completed flushing; nothing needed to be retained. Commitlog position was {}",
context.get());
}
return ssTable;
}
catch (Throwable e)
{
writer.abort();
throw Throwables.propagate(e);
}
}