Overview
Since upgrading to PGO 6.0 and pg18, I've had to switch to OpenTelemetry metrics because the old Exporter is no longer available. It works fine except that I've seen a very steep increase in AWS S3 data egress fees from the pgbackrest buckets I'm using for backup. It's not proportionate to the size of the dataset being backed-up: a very small database is costing around as much in egress as a much larger one. Removing the instrumentation: {} spec stops metrics and the egress fees drop back down to the expected near-zero level.
It's too expensive to sustain, so I've had to disable metrics on the pg18 clusters until this problem can be addressed.
Perhaps the instrumentation: {} config in a PostgresCluster spec could be expanded to allow the user to set a level of backup monitoring which works with the constraints of their egress budget?
Environment
Please provide the following details:
- Platform:
Kubernetes
- Platform Version:
1.35.3
- PGO Image Tag:
ubi9-6.0.2-0
- Postgres Version:
18
- Storage: AWS S3
Steps to Reproduce
REPRO
Provide steps to get to the error condition:
- Add
instrumentation: {} to a PostgresCluster spec which uses postgres 18 and a pgbackrest backup to an AWS S3 bucket.
- Ensure the PGO has the
OpenTelemetryMetrics=true feature flag set.
- Monitor AWS egress cost.
EXPECTED
- For data egress and request fees to run at near-zero cost.
ACTUAL
- Egress fees from buckets run at several dollars per bucket, per day, which adds up to a significant monthly bill over multiple buckets.
Logs
The jump in July corresponds with rollout to a subset of our staging pg clusters:
Note the lack of variation in the cost of storage over the same period.
Additional Information
I can provide any more detail you require on request in a non-public channel.
Overview
Since upgrading to PGO 6.0 and pg18, I've had to switch to OpenTelemetry metrics because the old Exporter is no longer available. It works fine except that I've seen a very steep increase in AWS S3 data egress fees from the pgbackrest buckets I'm using for backup. It's not proportionate to the size of the dataset being backed-up: a very small database is costing around as much in egress as a much larger one. Removing the
instrumentation: {}spec stops metrics and the egress fees drop back down to the expected near-zero level.It's too expensive to sustain, so I've had to disable metrics on the pg18 clusters until this problem can be addressed.
Perhaps the
instrumentation: {}config in aPostgresClusterspec could be expanded to allow the user to set a level of backup monitoring which works with the constraints of their egress budget?Environment
Please provide the following details:
Kubernetes1.35.3ubi9-6.0.2-018Steps to Reproduce
REPRO
Provide steps to get to the error condition:
instrumentation: {}to aPostgresClusterspec which uses postgres 18 and a pgbackrest backup to an AWS S3 bucket.OpenTelemetryMetrics=truefeature flag set.EXPECTED
ACTUAL
Logs
The jump in July corresponds with rollout to a subset of our staging pg clusters:
Note the lack of variation in the cost of storage over the same period.
Additional Information
I can provide any more detail you require on request in a non-public channel.