When appending to the head and a chunk is full it is flushed to the disk and m-mapped (memory mapped) to free up memory
Prom startup now happens in these stages
- Iterate the m-maped chunks from disk and keep a map of series reference to its slice of mmapped chunks.
- Iterate the WAL as usual. Whenever we create a new series, look for it's mmapped chunks in the map created before and add it to that series.
If a head chunk is corrupted the currpted one and all chunks after that are deleted and the data after the corruption is recovered from the existing WAL which means that a corruption in m-mapped files results in NO data loss.
[Mmaped chunks format](https://github.com/prometheus/prometheus/blob/master/tsdb/docs/format/head_chunks.md) - main difference is that the chunk for mmaping now also includes series reference because there is no index for mapping series to chunks.
[The block chunks](https://github.com/prometheus/prometheus/blob/master/tsdb/docs/format/chunks.md) are accessed from the index which includes the offsets for the chunks in the chunks file - example - chunks of series ID have offsets 200, 500 etc in the chunk files.
In case of mmaped chunks, the offsets are stored in memory and accessed from that. During WAL replay, these offsets are restored by iterating all m-mapped chunks as stated above by matching the series id present in the chunk header and offset of that chunk in that file.
**Prombench results**
_WAL Replay_
1h Wal reply time
30% less wal reply time - 4m31 vs 3m36
2h Wal reply time
20% less wal reply time - 8m16 vs 7m
_Memory During WAL Replay_
High Churn:
10-15% less RAM - 32gb vs 28gb
20% less RAM after compaction 34gb vs 27gb
No Churn:
20-30% less RAM - 23gb vs 18gb
40% less RAM after compaction 32.5gb vs 20gb
Screenshots are in [this comment](https://github.com/prometheus/prometheus/pull/6679#issuecomment-621678932)
Signed-off-by: Ganesh Vernekar <cs15btech11018@iith.ac.in>
