Add documentation about performing manual queries on the DB

Author: Kobzol

database/queries.md

@@ -0,0 +1,58 @@
+# Useful queries
+This document contains useful queries that should be performed manually in exceptional situations.
+
+## Remove data for an artifact from the DB
+This is important for situations where there is some compilation error for a stable benchmark,
+with a stable release of the compiler. While this should be rare, it happens sometimes e.g. because
+of future incompatibility lints turning into errors.
+
+The set of queries below show an example of removing the measured data and errors for the `html5ever`
+benchmark for Rust `1.69` and `1.70`.
+```sql
+-- Remove 1.69 and 1.70 html5ever data and errors
+-- Delete errors
+DELETE FROM error
+WHERE aid IN (
+    SELECT id
+    FROM artifact
+    WHERE name IN ('1.69.0', '1.70.0') AND
+          type = 'release'
+);
+
+-- Delete collection records. This will enable `bench_published` to re-run the benchmarks.
+DELETE FROM collector_progress
+WHERE aid IN (
+    SELECT id
+    FROM artifact
+    WHERE name IN ('1.69.0', '1.70.0') AND
+          type = 'release'
+) AND step IN ('html5ever');
+
+-- Note that there probably won't be any results in pstat, since the benchmark did not compile.
+-- But it's possible that some configurations of it have compiled before. To avoid skewing
+-- the average, we should thus still delete.
+
+-- Delete compile benchmark data
+DELETE FROM pstat
+WHERE aid IN (
+    SELECT id
+    FROM artifact
+    WHERE name IN ('1.69.0', '1.70.0') AND
+          type = 'release'
+);
+
+-- Delete runtime benchmark data
+DELETE FROM runtime_pstat
+WHERE aid IN (
+    SELECT id
+    FROM artifact
+    WHERE name IN ('1.69.0', '1.70.0') AND
+          type = 'release'
+);
+```
+
+After running these queries, the benchmark should be fixed so that it is compilable again, and then
+the version should be re-benchmarked:
+```bash
+$ cargo run --bin collector bench_published <version>
+```