You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
As mentioned in the https://github.com/scylladb/scylla-enterprise/issues/4006, it might be beneficial to always send more token ranges per job than max_repair_ranges_in_parallel (e.g. the greater of 10 * max_repair_ranges_in_parallel and 5%-10% of all ranges owned by repaired replica set). It allows for better concurrency (the whole job won't be blocked from returning when only 1 token range is still being repaired). Note that this should be safe to do with non-max intensity as well. It would require SM to use the ranges_parallelism repair param that limits used parallelism even when SM sends less ranges than max_repair_ranges_in_parallel.
This behavior could be controlled by an additional flag or repair config option in scylla-manager.yaml.
A side benefit would be to decrease the amount of clutter in SM logs.
In terms of testing, it would be good to see performance improvement on a cluster like: 2ds, 5 nodes each, keyspace with RF 3 in each dc, setup in which the repair indeed has to do some work (missing rows on some nodes).
The text was updated successfully, but these errors were encountered:
As mentioned in the https://github.com/scylladb/scylla-enterprise/issues/4006, it might be beneficial to always send more token ranges per job than
max_repair_ranges_in_parallel
(e.g. the greater of10 * max_repair_ranges_in_parallel
and5%-10%
of all ranges owned by repaired replica set). It allows for better concurrency (the whole job won't be blocked from returning when only 1 token range is still being repaired). Note that this should be safe to do with non-max intensity as well. It would require SM to use theranges_parallelism
repair param that limits used parallelism even when SM sends less ranges thanmax_repair_ranges_in_parallel
.This behavior could be controlled by an additional flag or repair config option in
scylla-manager.yaml
.A side benefit would be to decrease the amount of clutter in SM logs.
In terms of testing, it would be good to see performance improvement on a cluster like: 2ds, 5 nodes each, keyspace with RF 3 in each dc, setup in which the repair indeed has to do some work (missing rows on some nodes).
The text was updated successfully, but these errors were encountered: