Background: We are using a job using ProcessFunction which reads data from
kafka fires ~5-10K timers per second and sends matched events to KafkaSink.
We are collecting metrics for collecting no of active timers, no of timers
scheduled etc. We use statsd reporter and monitor using Grafana dashboard &
RocksDBStateBackend backed by HDFS as state.
1. *Counter value suddenly got reset:* While job was running fine, on one
fine moment, metric of a monotonically increasing counter (Counter where we
just used inc() operation) suddenly became 0 and then resumed from there
onwards. Only exception in the logs were related to transient connectivity
issues to datanodes. Also there was no other indicator of any failure
observed after inspecting system metrics/checkpoint metrics. It happened
just once across multiple runs of a same job.
2. *Counters not retained during flink restart with savepoint*: Cancelled
job with -s option taking savepoint and then restarted the job using the
savepoint. After restart metrics started from 0. I was expecting metric
value of a given operator would also be part of state.
3. *Counter metrics getting sent as Gauge*: Using tcpdump I was inspecting
the format in which metric are sent to statsd. I observed that even the
metric which in my code were counters, were sent as gauges. I didn't get why
that was so.
Can anyone please add more insights into why above mentioned behaviors would
Also does flink store metric values as a part of state for stateful
operators? Is there any way to configure that?
View this message in context:
Sent from the Apache Flink User Mailing List archive. mailing list archive at