jonnify·bcs /bcs/elixir-core/champ

B2.3 · The CHAMP Property Database — manuscript chapter 2.3

A snapshot is a structure, not a copy.

Chapter 2.2's stores own the present perfectly; the desk asks questions in other tenses — what did the book look like at 14:30, show me the portfolio as-of yesterday's close, fork this risk scenario and mutate the copy. The CHAMP forest is the BEAM-native answer: structural sharing as the snapshot mechanism, the contract hash as the trie's placement function — the part preface's guideline, cashed here with words counted.

The chapter is content/bcs2.3.md. The implementation is the Operator's production drop — three modules, 1,458 lines, delivered on June 7 — integrated under review with one gated amendment; the rung behind it is bcs2.3, and its committed transcript closes PASS 7/7.

The seven gates, mapped to the dives. Drawn in transcript order. Select a gate to read its verbatim line and the dive that teaches it.
H1 canon dive 1 H2 persist dive 2 H3 sharing dive 2 H4 forest dive 3 H5 hotpath dive 3 H6 crossover dive 3 H7 server dive 3
readout · Select a gate to read its verbatim line and the dive that teaches it. Seven gates over the integrated drop; the full transcript is frozen in §2 below, the dives are mapped in §3.
§1

The tenses the flat table cannot serve

content/bcs2.3.md · Why

The moment history or hypotheticals become product features, every answer from the flat table is a snapshot, and a snapshot of an ETS table is a copy-out priced per row. Structural sharing is the BEAM-native alternative — the immutable heap was built for exactly this — and this chapter cashes the guideline with words counted.

THE DROP

Integrated under review

The production-drop pattern returns in its original direction: three modules, 1,458 lines, delivered on June 7 — and the review earned its keep.

THE AMENDMENT

One placement law, in-process too

The drop's placement hash was a one-round mix; it now delegates to the contract's hash32 — three lines and a comment, gated by H1.

THE HONEST METRIC

Marginal cost, not a percentage

The first sharing metric capped near fifty and failed a correct implementation; marginal cost is the metric of record — gated by H2 and H3.

THE CROSSOVER

Both directions, measured

The flat table wins the present; the forest wins history categorically — and the chronology trade is named on stage. Gated by H4–H7.

§2

The proof

content/bcs2.3.md · quoting bcs_rung_2_3_check.out

The full committed transcript, verbatim — the [debug] boot line stays because the drop logs its starts, and the record keeps the drop's voice:

content/echo_data/runtimes/elixir/bcs_rung_2_3_check.out · verbatimfrozen
H1 canon ok -- placement delegated to the contract hash: compute_hash_int -> BrandedId.hash32; reference snowflake round-trips
H2 persist ok -- the old snapshot is intact: v1 holds 1000, v2 holds 1001, and v1 cannot see the new row
H3 sharing ok -- structural sharing measured: v2 costs 122 words beside v1 (one path copy) against 6688 standalone -- 98% of v2 is shared with v1
H4 forest ok -- two kinds, one forest: namespace_size AST=500 ORD=500 from the cached census, partition exact
H5 hotpath ok -- by-snowflake 1858 ns/op vs string-id 2344 ns/op over 20000 puts
H6 crossover ok -- at 20000 rows -- build: champ 41 ms vs ets 7 ms; point read: champ 799 ns vs ets 315 ns; snapshot: 1000 champ snapshots in 90 us (a snapshot is a binding) vs one ets copy-out in 2540 us

07:31:13.583 [debug] ChampServer started with 0 entries
H7 server ok -- read scale-out: 10000 server calls 28 ms vs snapshot-once-then-pure 8 ms (copy cost included)
PASS 7/7

The integrated modules live at runtimes/elixir/lib/echo_data/champ/ — branded_champ.ex, champ_node.ex, champ_server.ex — with the drop's provenance and the single amendment recorded in the source comment. The documented surface: put/3, fetch/2, the *_by_snowflake family, get_namespace/2, namespace_size/2, get_champ/1.

Boundaries, as the manuscript states them: writes serialize through the owning process; chronological reads over the forest are a sort or a sibling index; no disk persistence — the durability story remains where Part II put it; GC interplay under heavy version churn is real and unmeasured here. The BCS-grade shell — namespace-scoped, Bcs.gate at every call, typed errors — is the carried integration rung, not a patch smuggled into the drop.

§3

The dives

three lessons, fixed by the chapter spec
B2.3.1

The Forest and the Placement Law

H1 — the namespace→trie forest keyed by the snowflake integer; compute_hash_int -> BrandedId.hash32 — one placement law, in-process too; the two-bitmap CHAMP node: datamap/nodemap, popcount-indexed, 5-bit fragments.

dive 1built B2.3.2

Sharing at the Honest Metric

H2 — v1 holds 1000, v2 holds 1001, the new row absent from v1. H3 — 122 words beside v1, 98% shared, against 6688 standalone; the denominator trap, documented so it is paid for once.

dive 2built B2.3.3

The Crossover

H4–H7 — the forest census, the by-snowflake hot path, the crossover stated plainly in both directions, chronology lost to hash order, the snapshot-out pattern, the tense litmus.

dive 3built

Part II continues past this module — Archetypes and Composition (B2.4) composes definitions at read time by a pure fold, and Relations Are Systems (B2.5) promotes edges to systems.

§4

References

sources · related

Sources

Related