Summary: Can now replay a full tuning run from Proto-X. This is used to see how each step of tuning would have done without query timeouts and without Boot enabled.
Demo:
The image shows the data of a replayed run of TPC-H SF0.01 without Boot enabled during tuning. For each step of tuning, the replay shows the # of queries executed during the original run (which may be < 22 if the workload timed out), the # of queries that timed out during the original run, and whether the workload timed out. It also shows this same information about the replay. You can see that the replayed times are always >= the original times, which makes sense. Whenever the original run is "22,0,False" (i.e. 22 executed, 0 timed out, workload didn't time out), the replayed time matches closely.
Details:
Migrated the pipeline which pickles actions (DBMS configuration changes) while tuning and replays actions during replay to get the DBMS into the correct state.
Related to the above, modified the pipeline to store both the best per-query knobs found during execute_variations() as well as all per-query knob variations tried. We can either replay the best variation or all variations. This is especially useful if the workload timed out in the original run, in which case the "best" variation is a misnomer as it is simply an arbitrary variation.
Made Proto-X log additional information while tuning about the # of executed queries, # of timed out queries, and whether the workload timed out. During replay, all this information may now be utilized.
Fixed a bug where reset() was overwriting the logged replay information for a step, leading to a mismatch between the dumped action.pkl file (which contains the DBMS configuration state) and the run.raw.csv file (which contains the runtime information of the workload during that state).
Refactored all symlinks to have the .link extension to fix a subtle bug where a replay would overwrite the output.log file of the original run.
Standardized whether the page cache is dumped during tuning and replay (it's now not dumped in either case).
Made CLI options for the time to run the RL agent and whether Boot is enabled more fine-grained such that these values can differ between HPO and tune.
Summary: Can now replay a full tuning run from Proto-X. This is used to see how each step of tuning would have done without query timeouts and without Boot enabled.
Demo: The image shows the data of a replayed run of TPC-H SF0.01 without Boot enabled during tuning. For each step of tuning, the replay shows the # of queries executed during the original run (which may be < 22 if the workload timed out), the # of queries that timed out during the original run, and whether the workload timed out. It also shows this same information about the replay. You can see that the replayed times are always >= the original times, which makes sense. Whenever the original run is "22,0,False" (i.e. 22 executed, 0 timed out, workload didn't time out), the replayed time matches closely.
Details:
execute_variations()as well as all per-query knob variations tried. We can either replay the best variation or all variations. This is especially useful if the workload timed out in the original run, in which case the "best" variation is a misnomer as it is simply an arbitrary variation.reset()was overwriting the logged replay information for a step, leading to a mismatch between the dumpedaction.pklfile (which contains the DBMS configuration state) and therun.raw.csvfile (which contains the runtime information of the workload during that state)..linkextension to fix a subtle bug where a replay would overwrite theoutput.logfile of the original run.