Observability
Nine tools, each answering a different question about a running program. Find your question in the left column.
| Your question | What answers it |
|---|---|
| Why is this answer set empty? | rows.why(), one sentence naming the pattern, join, or guard that killed it |
| How was this answer derived? | metta.derivation: Derivation, Step, Fact proof trees |
| What will this query do, before I run it? | prepare(...).explain() and cursor.explain(), the plan reflected |
| What did this evaluation do, step by step? | metta.trace, the reduction trace as events |
| What did this call cost? | m.stats(), engine counter deltas over a with block |
| Where did the time go? | m.profile(), the engine's own profiler over a block |
| What is tabling holding? | (table-stats): tables, answers, hits, invalidations |
| What is silently wrong in this space? | metta.lint; lint_file(path) anchors each finding to its file:line |
| What is changing, as it changes? | m.subscribe(pattern), a standing query over writes |
Three that save the most time
Ask why() before adding prints. When a query returns nothing, why() already knows which conjunct produced no rows. Re-running with print statements finds out the same thing more slowly.
Read explain() before profiling a slow foreign query. The usual cause is a pattern that stopped pushing down into the backend, and explain() shows that without running the query at all.
Trust stats() inferences over wall clock. The inference counter is deterministic: the same workload gives the same number on any machine, under any load. Wall clock does not. This repository gates its own benchmarks on inferences for that reason.
The engine describes itself
What the engine has registered, declared, and served is stored as ordinary atoms in the &metta space, so you can query it the same way you query anything else. Reflection and steering covers that.