Photon event logs do not store certain metrics, such as scan time, shuffle write time, and peak execution memory, in the same format as CPU Spark event logs. These metrics are used by QualX for prediction purposes.
Missing Metrics/Features
Feature
Type
scan_time
Spark Metric
sw_writeTime_mean
Spark Metric
peakExecutionMemory_max
Spark Metric
sqlOp_SubqueryBroadcast
Exec
sqlOp_RunningWindowFunction
Exec
sqlOp_Expand
Exec
Solution
After investigation we found alternative ways to calculate some of these metrics:
PhotonScan nodes provide a cumulative time metric that can be used as a replacement for the scan time metric.
shuffle write time can be reconstructed using the following metrics:
time taken waiting on file write IO (part of shuffle file write)
time taken to sort rows by partition ID (part of shuffle file write)
time taken to convert columns to rows (part of shuffle file write)
Photon nodes provide a peak memory usage metric, which can be used for the peak execution memory metric.
Describe the bug
Photon event logs do not store certain metrics, such as scan time, shuffle write time, and peak execution memory, in the same format as CPU Spark event logs. These metrics are used by QualX for prediction purposes.
Missing Metrics/Features
Solution
After investigation we found alternative ways to calculate some of these metrics:
PhotonScan
nodes provide acumulative time
metric that can be used as a replacement for thescan time
metric.shuffle write time
can be reconstructed using the following metrics:peak memory usage
metric, which can be used for thepeak execution memory
metric.cc: @amahussein @leewyang