|
| 1 | +# Σ.SP — Spillable-join arc: grace-partitioned join demotion |
| 2 | + |
| 3 | +Status: **Phase 0 (design + box re-baseline), 2026-07-11.** Successor to |
| 4 | +the "paging arc" in [MEMORY_BUDGET.md](MEMORY_BUDGET.md) — that doc's |
| 5 | +decision gate said *reassess after the next DataFusion upgrade*; the |
| 6 | +reassessment is done and recorded here. |
| 7 | + |
| 8 | +## Why now, and what changed since MEMORY_BUDGET.md |
| 9 | + |
| 10 | +The paging arc was framed when Q09 thrashed and parted-SF100 tar-pitted. |
| 11 | +Both of those are gone by other means: |
| 12 | + |
| 13 | +- **Q09 is solved at the plan level** (Σ.JS.1/Σ.JS.2, PRs #176/#178): |
| 14 | + honest join-side estimates keep builds small; SF=100 flat suite 51.8 s. |
| 15 | +- **The parted tar pit was scan-width, not paging** (Σ.MW.1, PR #179): |
| 16 | + the multi-file union multiplied decode width by the part count. With |
| 17 | + the budget split, the full parted single-node SF=100 suite completes |
| 18 | + (47.7 s / 22.4 GB peak on the 36 GB dev box — first complete run |
| 19 | + anywhere; 32 GB box validation pending, Phase 0). |
| 20 | + |
| 21 | +What remains for THIS arc is the last structural weakness: **oversized |
| 22 | +tracked join builds cannot yield memory.** DF 53's `HashJoinExec` cannot |
| 23 | +spill; under a cap it deadlocks (the refuted 0.7×RAM blanket cap), and |
| 24 | +against the `ElasticFloorPool` it fails the query cleanly at best, |
| 25 | +kernel-OOMs at worst when the growth is bursty. Today's exposure: |
| 26 | + |
| 27 | +- Q18/Q21-class plans (giant semi/anti/inner builds) ride the margin on |
| 28 | + 32 GB boxes: ~28 GB peaks that pass or die on page-cache luck. |
| 29 | +- Any hostile/ad-hoc workload (not TPC-H) can synthesize a build larger |
| 30 | + than RAM and there is no graceful degradation. |
| 31 | + |
| 32 | +## Upstream status (checked 2026-07-11) |
| 33 | + |
| 34 | +- Spillable hash join has **not landed through DataFusion 54.0.0** |
| 35 | + (June 2026, latest). The proposal epic is |
| 36 | + [apache/datafusion#17267](https://github.com/apache/datafusion/issues/17267) |
| 37 | + (hybrid hash join with greedy partition spilling) — design stage, |
| 38 | + "not yet fully underway". apache/datafusion#12952 and #1599 are the |
| 39 | + long-running asks. |
| 40 | +- Conclusion per the MEMORY_BUDGET decision gate: build in-repo, keep |
| 41 | + it a *demotion* (stock plans untouched unless the estimate says |
| 42 | + danger), and treat upstream as the eventual replacement. |
| 43 | + |
| 44 | +## Design: plan-time demotion, not runtime panic |
| 45 | + |
| 46 | +The upstream proposal spills *reactively* (notice memory pressure |
| 47 | +mid-build, spill largest partitions). We have something DF doesn't: the |
| 48 | +Σ.JS.2 **grounded bottom-up row/byte estimates** at plan time. That |
| 49 | +enables the simpler, more predictable half of the design space: |
| 50 | + |
| 51 | +**A physical rule (`GraceJoinDemotionRule`, after Σ.JS.1 in the preset |
| 52 | +chain) rewrites a `HashJoinExec` whose GROUNDED build-side estimate |
| 53 | +exceeds a budget into a `GraceHashJoinExec`.** |
| 54 | + |
| 55 | +- **Budget:** `EMAT_GRACE_BUILD_BYTES` tri-state. AUTO = demote when |
| 56 | + estimated build bytes > `f × sensed MemAvailable at plan time` |
| 57 | + (start f = 0.5; the Σ.AI.6c sensor already exists and is cached). |
| 58 | + Ungrounded estimates never demote (same honesty rule as Σ.JS.1 — |
| 59 | + no decision without proof), so TPC-H flat plans are untouched: |
| 60 | + bench == release means banked numbers cannot move. |
| 61 | +- **Operator (`GraceHashJoinExec`):** classic grace hash join. |
| 62 | + 1. Partition BOTH inputs to `K` spill files by `hash(join keys)` |
| 63 | + (reuse DF's IPC spill-file machinery used by sort/agg spills). |
| 64 | + 2. Join partition pairs sequentially with a stock in-memory |
| 65 | + `HashJoinExec` over per-partition record-batch streams; a pair |
| 66 | + whose build still exceeds the budget recurses with a different |
| 67 | + hash seed (bounded depth, then hard error). |
| 68 | + 3. `K` = ceil(estimated build bytes / (budget / 2)) rounded to a |
| 69 | + power of two, min 4. |
| 70 | +- **Join types, phased:** Inner first (Phase 1); LeftSemi/LeftAnti |
| 71 | + (the Q21 shape) in Phase 2; outer joins only if evidence demands. |
| 72 | +- **Memory accounting:** partition write buffers + the per-pair build |
| 73 | + reserve from the session pool, so the ElasticFloorPool sees the |
| 74 | + truth. Spill bytes/files surfaced as metrics |
| 75 | + (`grace_partitions`, `grace_spill_bytes`, `grace_recursions`). |
| 76 | + |
| 77 | +### Why demotion instead of a spilling retrofit of HashJoinExec |
| 78 | + |
| 79 | +- The refuted-cap evidence shows mid-build reactive spilling under |
| 80 | + DF 53's scheduling is deadlock-prone (concurrent reservations wait on |
| 81 | + each other). A demoted plan never enters that regime: partitioning is |
| 82 | + streaming with O(K buffers) memory, and each pair-join is sized to |
| 83 | + fit by construction. |
| 84 | +- SMJ demotion was already tried and refuted (MEMORY_BUDGET evidence |
| 85 | + table: SMJ+cap = `ResourcesExhausted`, SMJ unbounded = kernel OOM). |
| 86 | +- Plan-time demotion is testable deterministically (feed the rule a |
| 87 | + plan with grounded stats; assert the rewrite), matching the Σ.JS test |
| 88 | + discipline. |
| 89 | + |
| 90 | +## Phases |
| 91 | + |
| 92 | +- **Phase 0 — re-baseline (cheap, decisive).** One 32 GB box run of the |
| 93 | + post-Σ.MW.1 parted single-node suite. If Q18/Q21 now pass with the |
| 94 | + width fix, the arc's default posture can stay conservative (demotion |
| 95 | + fires only on genuinely-oversized builds, expected: none in TPC-H). |
| 96 | + Exit: banked baseline + updated exposure list. |
| 97 | +- **Phase 1 — Inner grace join behind `EMAT_GRACE_JOIN` (default OFF).** |
| 98 | + Operator + demotion rule + unit/property tests (result parity vs |
| 99 | + stock join across seeds/partition counts, spill-file cleanup, |
| 100 | + accounting) + a synthetic oversized-build workload that OOMs stock |
| 101 | + and completes demoted. |
| 102 | +- **Phase 2 — semi/anti support + full-suite A/B.** The 22q protocol on |
| 103 | + the 32 GB box, flat + parted, default-ON decision by the same rule as |
| 104 | + every memory lever: zero regression on healthy runs, graceful |
| 105 | + degradation on pressure runs. |
| 106 | +- **Phase 3 — upstream watch.** When apache/datafusion#17267 lands, |
| 107 | + benchmark theirs vs ours; prefer deleting ours. |
| 108 | + |
| 109 | +## Non-goals |
| 110 | + |
| 111 | +- Replacing the ElasticFloorPool (it stays as the last-resort guard). |
| 112 | +- Spilling aggregates/sorts (DF already does; multi-level merge since |
| 113 | + DF 50). |
| 114 | +- Symmetric/streaming joins. |
0 commit comments