-
Notifications
You must be signed in to change notification settings - Fork 38
Description
Currently, for YDB disks, all read/write requests go through the Partition tablet.
At the same time, the partition tablet performs synchronous read transactions from the local database during reads and asynchronous operations (Flush, Compression), during which it is blocked. That is, while the Execute transaction is running, the entire tablet is blocked and cannot process writes, even those that go to fresh and do not require writing to the index.
These transactions can take 5-10 ms to complete, as they require reads from the blobstorage.
Max Execution time of transactions on hosts and percentiles per all hosts:

Because of this, when Compaction begins, the latency doubles and it becomes as unstable as possible.
It is supposed to write a Proxy actor in front of the Partition tablet, which will write fresh blocks in Blob storage, but in order to read, write something to the index in the local database, it will go to the Partition tablet, started nearby in the external boot.