docs(sales-summaries): record the end-to-end run on the restore

Executes what Part 1 only measured: Phase 0, the re-key migration, a live Square
import, and a full recompute through sales-summaries-v2, with every window
re-cut to end 08-13 so the backup's partial final day stops skewing the numbers.

Reportable error falls from 542 days off / $30,982.11 to 71 days off / $405.79,
a 98.7% reduction, with 5 material days left — three of them the ones the plan
predicted. 107,815 entities were re-keyed with refund and charge counts
unchanged, which is what proves the expand step resolves legacy keys instead of
duplicating under the new scheme. NGCD 2026-07-23, off by exactly $71.94 with no
refund line, now carries the line and balances to 9.1e-13.

Records two defects that only executing could surface:

Re-keying a charge with two parent orders duplicates the tender. The migration
scoped such charges through whichever order it saw first, so the other client's
import found neither the scoped nor the legacy key and created a second charge,
which cardinality-many then added beside the first. Contended clients' error was
$441,045 until 15,961 stale refs were retracted. §4.3 warned about this and the
migration under-weighted it; it must split shared charges before shipping.

dirty-sales-summaries index-pulls from the client's start point with no :end and
filters lazily, so it scans every later client's summaries too — O(n²), and the
recompute degraded from ~180 to ~3 client-days per minute as summaries piled up.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
2026-08-15 08:10:32 -07:00
parent e24ffa3bdf
commit 8913f545f3

View File

@@ -168,3 +168,122 @@ It has `client` and `location` in scope already, so the fix is local.
- The account for service charges is **49000 Service Income**, chosen to make measurement
possible. It does not affect balance at all — only whether a day can reach `accepted?` — so
every number here is independent of that choice.
---
# Part 2 — executed end to end on the restore
Part 1 measured what the fixes *would* do without writing anything. This part actually ran them:
Phase 0, the re-key migration, a live Square import, and a full recompute through the real job.
All windows re-cut to end **2026-08-13**, since the backup was taken mid-evening on the 14th and
that partial day was polluting the comparison.
## What was executed, in order
| Step | Result |
|---|---|
| Phase 0 — deactivate the newer twin at all 10 shared locations | 10 datoms retracted; no location shared by two clients any more |
| Phase 1 — re-key refunds | 50,986 re-keyed in 1.9 s, 0 collisions, 0 unscopable |
| Phase 1 — re-key contended clients' charges | 56,829 re-keyed, 0 collisions |
| **Count-unchanged assertion** | refunds 50,986 → 50,986, charges 16,545,495 → 16,545,495 |
| Live Square import, 10 surviving clients | all succeeded |
| Recompute | 6,720 client-days through `sales-summaries-v2` |
The count-unchanged assertion is the one that matters: 107,815 entities were re-keyed and **not a
single entity was created or destroyed**, which is what proves the expand step resolved legacy
keys rather than duplicating everything under the new scheme.
## Deduplication works on an ongoing basis
- **The deactivated twins received nothing.** Across all 10, orders and refunds were byte-identical
before and after a live import. Phase 0 holds.
- **Refunds stopped alternating and became two stable copies.** 98 of NGCC's 125 refunds are now
also held by NGCD as separate entities. Each client owns its own; neither can take the other's.
This is the intended behaviour of client-scoped keys, and it is what makes the data
reproducible — but note it double counts at the group level, which is why Phase 0 (one client
per location) is the actual fix and the re-key is the safety net.
### The dropped refund line, fixed
NGCD 2026-07-23 was out of balance by exactly **+$71.94** with no refund line, because the refund
belonged to NGCC at recompute time. After the full pipeline the summary carries a **"Card Refunds"
line of $71.94** and the day's imbalance is **9.1e-13 — zero**.
## Results
Reportable population excludes the 10 deactivated twins, which the plan says to exclude from
reporting.
| | Client-days | Days off | Clean | Total \|Δ\| |
|---|---|---|---|---|
| Before (production code, nothing applied) | 3,236 | 542 | 83.25% | $30,982.11 |
| **After (everything applied)** | 3,030 | **71** | **97.66%** | **$405.79** |
A **98.7% reduction in dollar error**. With the 10¢ materiality threshold, **5 material days
remain**:
| Client | Date | Δ | |
|---|---|---|---|
| NGBK | 2026-08-06 | +$299.42 | new — see below |
| NGDA | 2026-08-01 | $50.00 | the known auto-gratuity-as-service-charge order |
| NGEB | 2026-08-10 | $25.00 | ezCater fee semantics, predicted |
| NGEB | 2026-07-29 | $20.00 | ezCater fee semantics, predicted |
| NGPS | 2026-08-12 | +$9.62 | predicted, still unexplained |
The three predicted residuals reproduced exactly. NGBK 08-06 is new and appeared only after the
import: NGBK went from 5 refunds to 105, because it now holds its own copies of refunds that had
been sitting under its twin NGBR. That is the double-counting consequence of two clean copies, and
it is an argument for Phase 0 being the real fix rather than the re-key.
## Two defects found by executing rather than simulating
### 1. Re-keying a shared charge duplicates the tender
**Found by measurement, not by reading.** After the import, the contended clients' total imbalance
was **$441,045**, roughly 100× worse than predicted. The cause:
1. 35,870 of 56,829 charges had **two parent orders**, one per twin.
2. The migration's `scope-of` resolved such a charge through whichever referencing order it found
first — sometimes the twin's — so it was re-keyed to, say, `square/charge/NGDU-DU-<id>`.
3. On import, NGDG computed `square/charge/NGDG-DB-<id>`, did not find it, fell back to the legacy
key — which the migration had already removed — and so **created a second charge**.
4. `:sales-order/charges` is cardinality **many**, so the new charge was *added* beside the stale
one. One order, two tenders for one payment. Tender roughly doubled.
Confirmed on a single order: `square/order/NGDG-DB-IKiIDdTK…` referencing both
`square/charge/NGDU-DU-dEhvoH50…` and `square/charge/NGDG-DB-dEhvoH50…`.
This is exactly what plan §4.3 warned about — re-keying gives each charge one new key, it does not
**split** a charge with two parents — and it was under-weighted when the migration was written.
**Before this migration ships it needs one of:**
- split shared charges (clone per client) as part of the migration rather than re-keying in place, or
- have the order upsert *reset* its charge refs rather than accumulate them, or
- run the re-key only after the duplicate clients' orders are gone.
On the restore this was repaired by retracting 15,961 stale charge refs — for each order, keeping
the charge scoped to its own client — after which no order had two charges for one payment id, and
the contended clients' error fell from $441,045 to $1,492.
### 2. `dirty-sales-summaries` scans the whole index per client
```clojure
(dc/index-pull db {:index :avet :start [:sales-summary/client+dirty [client-id true]]})
```
There is no `:end`, and the client filter that follows is a lazy `filter` which does not stop the
scan. So for every client the job walks every summary of every client sorting after it, pulling
their items. It is O(n²) in the number of summaries.
Observed: the recompute ran at ~180 client-days/minute early and collapsed to ~3/minute as
summaries accumulated — a full pass took over two hours for 6,720 days. Adding an `:end` bound at
the client boundary should fix it.
## Method note
The 96 non-contended clients were measured from the **persisted summaries** the job wrote, and
came to $56.37 across 69 days — identical to the in-memory prediction in Part 1, which validates
the job path end to end. The 20 contended clients were re-measured in memory after the charge
repair, because a second full pass through the job would have taken hours for the reason in defect
2. The two methods agree exactly where both were run.