So - we have three types of times
- execTime, this is time actually spent in the CPU
- servTime, this is time spent waiting for services (indexer for scans, KV for fetches)
- kernTime, this is time spent waiting for the kernel to find free cpu time to execute. This can be due to two causes: overloaded cpus, or a bottleneck upstream or downstream.
In the case of Filter, the only concern is kernTime: this means that the filter has processed everything quickly (the execTime is 18.9ms) but has had to wait 38s in between receiving the documents and sending them on.
The projections have waited even longer: 1m27s in kernTime.
What I would say is that either the n1ql node is overloaded, or the stream operator has had to wait on the client to send the data.
Useful links to understand monitoring and profiling: