pybind/mgr/pg_autoscaler: Use bytes_used for actual_raw_used

Problem

We realized that `store` is not
the correct value to represent `actual_raw_used`
when it comes to pool(s) with `compression` enabled.

https://github.com/ceph/ceph/pull/29986
was the PR that is the culprit of the issue, since
it simply changed `byte_used` to `store` just
because they want a per pool value of bytes_used
without factoring in replication. However, they
did not realized that in doing so also caused
pools with compression to inherit an incorrect
value for `actual_raw_used`.

This also caused an incorrect value for `capacity_ratio`
since the autoscaler scales PGs according to the
`capacity_ratio` of each pool. The existing issue
causes pool with compression to have higher `capacity_ratio`
where in reality the actual utilization is less than
non-compressed pools, assuming we perform I/O with the same
work load on each pool evenly.

Solution

Use `bytes_used` instead of `store` when fetching
for `actual_raw_used` and when calculating `pool_raw_used`
we `max(actual_raw_used, target_bytes * raw_used_rate)`

Fixes:

https://tracker.ceph.com/issues/54136

Signed-off-by: Kamoltat <ksirivad@redhat.com>
(cherry picked from commit 3d8ac80f61cd332b53aea1fa5799f8ccd3b01b66)

diff --git a/src/pybind/mgr/pg_autoscaler/module.py b/src/pybind/mgr/pg_autoscaler/module.py
index 13e1cff9e8838655cf151df676847a2fc9153d1c..305a3d58dd1ea2a96acbc035b83a1fd0469da960 100644 (file)
--- a/src/pybind/mgr/pg_autoscaler/module.py
+++ b/src/pybind/mgr/pg_autoscaler/module.py
@@ -561,7 +561,6 @@ class PgAutoscaler(MgrModule):
 
             raw_used_rate = osdmap.pool_raw_used_rate(pool_id)
 
-            pool_logical_used = pool_stats[pool_id]['stored']
             bias = p['options'].get('pg_autoscale_bias', 1.0)
             target_bytes = 0
             # ratio takes precedence if both are set
@@ -569,10 +568,10 @@ class PgAutoscaler(MgrModule):
                 target_bytes = p['options'].get('target_size_bytes', 0)
 
             # What proportion of space are we using?
-            actual_raw_used = pool_logical_used * raw_used_rate
+            actual_raw_used = pool_stats[pool_id]['bytes_used']
             actual_capacity_ratio = float(actual_raw_used) / capacity
 
-            pool_raw_used = max(pool_logical_used, target_bytes) * raw_used_rate
+            pool_raw_used = max(actual_raw_used, target_bytes * raw_used_rate)
             capacity_ratio = float(pool_raw_used) / capacity
 
             self.log.info("effective_target_ratio {0} {1} {2} {3}".format(
@@ -616,7 +615,7 @@ class PgAutoscaler(MgrModule):
                 'crush_root_id': root_id,
                 'pg_autoscale_mode': p['pg_autoscale_mode'],
                 'pg_num_target': p['pg_num_target'],
-                'logical_used': pool_logical_used,
+                'logical_used': float(actual_raw_used)/raw_used_rate,
                 'target_bytes': target_bytes,
                 'raw_used_rate': raw_used_rate,
                 'subtree_capacity': capacity,