I would test my pipeline using the direct runner, and it would happily run on a sample. When I ran it on the Dataflow runner, it'll run for a hour, then get to a stage that would crash like so:
java.util.concurrent.ExecutionException: java.lang.RuntimeException: Error received from SDK harness for instruction -224: execute failed: panic: reflect: Call using main.HistogramResult as type struct { Key string "json:"key""; Files []string "json:"files""; Histogram palette.ColorHistogram "json:"histogram,omitempty""; Palette []struct { R uint8; G uint8; B uint8; A uint8 } "json:"palette"" } goroutine 70 [running]:>
This was because I forgot to register my HistogramResult type.
It would be useful if the direct runner tried to marshal and unmarshal all types, to help expose issues like this earlier.
Also, when running on Dataflow, the value of flags, and captured variables, would be the empty/default value. It would be good if direct also caused this behaviour. For example:
prefix := “X”
s = s.Scope(“Prefix ” + prefix)
c = beam.ParDo(s, func(value string) string {
return
prefix + value
}, c)
Will work prefix "X" on the Direct runner, but will prefix "" on Dataflow. Subtle behaviour, but I suspect the direct runner could expose this if it marshalled and unmarshalled the func like the dataflow runner.
Imported from Jira [BEAM-6372](https://issues.apache.org/jira/browse/BEAM-6372). Original Jira may contain additional context.
Reported by: bramp.
I would test my pipeline using the direct runner, and it would happily run on a sample. When I ran it on the Dataflow runner, it'll run for a hour, then get to a stage that would crash like so:
prefix := “X”
s = s.Scope(“Prefix ” + prefix)
c = beam.ParDo(s, func(value string) string {
return
prefix + value
}, c)