Materialize only what queries need

Why this skill

Definitions are about logical identity. Storage is about physical access cost. These are independent. When the spec says

Behavior = ⟨ExecutionSubgraph, Effects, ExitPoints⟩

it tells you what characterizes a Behavior — not how to represent it on disk. Reading the tuple as "store these three pieces" is a category mistake. The result is a representation that contains data no query asks for, at the cost of OOM-grade memory pressure on real-scale inputs.

This is the meta-pattern behind a class of "performance bugs" that resist patching: each patch makes the implementation slightly faster, but the fundamental shape of the persisted data is wrong by orders of magnitude.

In-action triggers (catch the wrong default before writing code)

Trigger	Reflect	Quick check
Naming a collection subgraph / members / contains / comprises on a parent entity	List queries that need membership	If none — drop the collection, store an aggregate
Spec says X = ⟨A, B, C⟩ and you're modeling A, B, C as fields/edges of X	Distinguish definition from storage	Definition can stay; storage ≠ definition
Estimated row/edge count = entities × average-children = millions	Cost-model the queries you'd serve	Most queries are answerable from a 64-byte hash
"Materialize once, query fast" intuition	Only true if storage_cost ≪ Σ query_cost. Usually false for forward closures over large graphs	Estimate aggregate query frequency × on-demand cost vs storage cost
2nd patch at "memory issue" hasn't fixed it	Climb from process to premise: am I storing the right thing at all?	Apply this skill; the leak is shape, not retention

On-action diagnostic: am I about to over-materialize?

For every persistent collection / edge type / table you're about to create, walk this list. Skip → don't materialize.

1. Name a concrete query — user-facing or internal — that needs THIS data in THIS shape. If you can't, stop.
2. Aggregate? Could a hash, count, max/p90, or Set<dimension> answer instead?
3. On-demand viable? Can the query be computed from existing edges in ≤ 1 s on the realistic-scale graph? If yes, don't precompute.
4. Direction inverted? Many "X contains Y" queries are answered better by "Y is reachable from X-entry" — runtime backward walk, not stored forward edges.
5. Frequency? A query that runs once per CI gate doesn't justify precompute. A query in a hot path that scales with graph size might.

If ALL queries answer "aggregate suffices" or "on-demand is fine" — drop the materialized collection.

Diagnostic when memory issues resist patching

Symptoms that point at over-materialization (vs. real leak):

• Memory grows roughly linearly with the number of parent entities being iterated, not with global state size.
• Targeted patches (chunked flush, explicit GC, freeing intermediate vars) reduce the slope but don't eliminate it.
• "Synthetic small fixture passes; real-scale OOMs" — synthetic tests don't show the explosion because per-parent payload is small.
• Time-to-OOM is roughly proportional to heap_cap. Doubling the cap doubles the time before OOM, instead of completing.

When 2-3 of these are true, the issue is storage shape, not allocation discipline.

Refactor recipe

When you've over-materialized:

1. Re-derive query list for the artifact under suspicion.
2. For each query, pick minimum storage:
   • Aggregate (hash, count, p90) → field on parent entity
   • On-demand traversal → no storage; document the query function instead
   • Repeated forward joins on stored data → maybe materialize, with explicit cost model in the spec
3. Drop the unneeded edges/rows. Existing analyzers/enrichers stop emitting them.
4. Add a Datalog rule or query function for queries previously served by the materialized form.
5. Update the spec / research doc: change X = ⟨A, B, C⟩ to clarify "A, B, C characterize X; storage = {hash, summary, ...}".

Anti-patterns

• "The research doc says it's a tuple, so I store all three components" — category error. Spec gives the concept; storage is independent.
• "Materialize once, query fast" without a cost model — usually false for forward-closure-over-large-graph cases.
• Synthetic-only verification — "200 small entities pass, ship it." Always verify on realistic-scale fixture early. The shape bug only shows up when per-parent payload is large.
• Patching the leak instead of the shape — chunked flush, explicit GC, WeakRef, etc. each look reasonable in isolation but accumulate without fixing the issue. After 2 failed patches, climb to premise.
• "It's fine for typical case" as exit excuse — if the failing case is a moderate codebase, real worst case is bigger, not smaller. Don't rationalize away the failure.

Concrete failure (Grafema, 2026-04-26)

Spec (research doc):

Behavior  = ⟨ ExecutionSubgraph, Effects, ExitPoints ⟩
ExecutionSubgraph = reachable(EntryPoint, CALLS ∪ READS_FROM) ∩ ...

Initial plan emitted BEHAVIOR -COMPRISES-> code_node edges per touched node. For 149 features × ~thousands of touched nodes per feature (real codebase), that's millions of edges materialized in RFDB.

Real user queries:

• Auto-doc → CONTRACT (params/returns) + effect summary. Not subgraph.
• Dedup → hash(sorted(transitive_call_targets)). Not subgraph.
• Cognitive metrics → counts + depth p90. Not membership.
• "What features does this code serve?" → backward callers traversal at query time. Not forward subgraph.

Three patch rounds at "memory leak" (chunked flush → trim retained → shared index cache) each cut growth rate but never eliminated it. Round 4 (premise reflection triggered by user pushback) revealed: no query needs the materialized edges. Drop them — peak memory falls ~1000×.

Connection to companion skills

• reflection-in-and-on-action: catches the trigger ("2nd patch attempt"). This skill provides the resolution at premise level.
• profile-before-rewrite: don't optimize the wrong thing. This skill is upstream — pick the right shape, then profile. If you've already profiled and the issue is "we're storing X bytes per entity", this skill says "consider not storing X at all".
• validate-assumptions-before-implementing: same family. Validate the premise (do we even need this data?) before validating the implementation (does the leak fix work?).

When NOT to apply

• Spec's tuple components are all small and bounded (e.g., a few fields).
• The query that needs full materialization runs frequently and is in a hot path.
• You've already verified at realistic scale that materialization is safe and the spec query mix justifies it.

Skill is about catching the default mistake of materializing because the spec describes the concept that way. Some materializations are correct; the discipline is to derive them from queries, not from definitions.