Experiment

Input for an Experiment Run

• Project name (ID)
• Server hostname + database name
• DB snapshot version/timestamp
• Postgres version
• EC2 instance type
• Change ("no change" or something changed, e.g. postgresql.conf or create index ...)
• Queries (two options: (a) pgreplay bin file, or (b) custom queries)

Change and Queries are specified when Experiment is created.

Optimization idea: based on all input options we can calculate "experiment run hash" to allow skipping "no change" runs.

Output of an Experiment Run

• Postgres log (goes to S3)
• JSON produced by pgBadger (goes to S3)
• Main values from pgBadger JSON (we post them to Postgres.ai metadata DB, using API)
• [todo] pg_stat_statements snapshots
• [todo] snapshots of other pg_stat_*** views

Experiment

Each experiment consists of at least two Experiment Runs. One of these Runs must be with Change="no change". All Experiment Runs for an Experiment must use the same type of EC2 instances.