Stux
MVP
- Joined
- Jun 2, 2016
- Messages
- 4,419
Sorta.
All writes are buffered up into a transaction group. When the transaction group is full or when the transaction timeout expires, the writes are all written to disk, and pretty much as efficiently as possible. And the next transaction group begins to fill.
Think 2-5 seconds
If the next txg fills before the last one flushes, the fs blocks.
Transactions are not erased, they stick around in ARC for a bit incase someone tries to read them. If they get read they stick around around longer, etc.
Sync writes are additionally written to slog or to zil on the primary array. Never to be read from again, unless the worst happens... to replay the lost sync transactions. Sync writes won't complete until the slog write finishes. Async writes finish as soon as the write completes to the txg in ram.
