On this one-host study, MongoDB fits six of eight compared fault scenarios on the selected timing metric: four on write outage and two on acknowledgement gap. PostgreSQL fits the two peer-only partition scenarios through the feed gate. There is no overall winner. MongoDB lost no acknowledged writes in the 20 peer-only partition cells, but its feed duplicated one effect in 3. PostgreSQL's strict variants also have an unresolved durability-observer anomaly, and five PostgreSQL files may overlap heavy work. These limits travel with the result.

The question was whether a database's primary could fail while an agent workload wrote proof records and a feed turned them into marks. This is Lab 03 on this site, EXP 021. The plan and earlier amendments were published before this result; amendment 02 was published after counted runs had started, and says so. Counted cells ran 8–10 November 2026. The result is internally evidenced: source cells and runner remain private, so readers can inspect the published records and identities but cannot independently rerun the study.

What ran

MongoDB 8.0.12 used three data-bearing replica-set members, majority journaled writes and PyMongo 4.10.1. PostgreSQL 16.10 used three nodes with Patroni 4.0.6, etcd 3.5.21 and psycopg 3.2.3/libpq multi-host. PostgreSQL ran both a commit-order-safe outbox and a permanent-slot pgoutput feed; MongoDB used a change stream. Effect and cursor were written in one transaction on both arms.

All nodes ran on one Windows/WSL2 Docker host (Docker Desktop 4.56.0; 6 Docker CPUs, about 9.7GiB memory). Killing a container simulated node death (F1). Disconnecting both networks simulated full isolation (F2a); disconnecting only peers left the client able to reach the old primary (F2b). F3 was a planned switchover. Process death (F1b), the Patroni timing floor and an outbox-only block are labelled context, never deciding compared verdicts.

There were 334 counted cells: 320 fault and 14 steady cells, all recorded valid. Twenty smoke cells never count; seven invalid attempts (five smoke, two counted) were retained and rerun. Each compared scenario has five paired repetitions on seed 1 and five on seed 2. The measured number is the median over eight writer threads: fault-to-first-ack outage for F1/F2, and longest acknowledgement gap ending after the switchover for F3.

The frozen rule and the result

The gate requires no lost acknowledged writes, no phantom writes, a complete once-only feed, and passing steady cells. PostgreSQL's passing variant is selected by its lowest seed 1 median of repetition medians. After both gates pass, at least four of five seed 1 pairs must be at least 20% shorter, and the seed 2 paired difference must have the same sign. If only one arm passes, it fits by the gate alone. Seed1 selection and comparison share data; seed 2 outage values are out of sample but its gate outcomes participate in selection.

F3 compares acknowledgement gaps; other rows compare write outage. A negative reproduction difference favours MongoDB. Gate-only rows do not compare outage.
ScenarioChosen PostgreSQL variantSeed1 wins, MongoDB / PostgreSQLSeed2 median difference, secondsVerdict and basis
default/F1classic-strict5 / 0-24.9642mongodb-fits · outage
default/F2aquorum-strict5 / 0-9.7462mongodb-fits · outage
default/F2bquorum-strict——postgres-fits · feed gate only
default/F3classic-strict5 / 0-3.1118mongodb-fits · ack gap
tuned/F1quorum-strict5 / 0-10.9708mongodb-fits · outage
tuned/F2aclassic-strict5 / 0-10.2184mongodb-fits · outage
tuned/F2bquorum-strict——postgres-fits · feed gate only
tuned/F3quorum-strict5 / 0-9.0009mongodb-fits · ack gap

F2b is a feed result, not lost MongoDB writes. The change stream duplicated one effect in 2 of 10 default cells and 1 of 10 tuned cells. All 20 passed the write gate: zero lost, phantom or duplicate-first writes. Every F2b cell had 7–9 accepted-later-rewritten writes, including the 17 whose feed passed; those rewrites are not lost acknowledged writes. A write-concern timeout and resume appear in failing and passing cells, so the mechanism is not established. None of the three failing cells is a rerun of a discarded cell. MongoDB's outage was shorter in all 20 pairs, but the frozen gate prevents an outage verdict here.

Components: PostgreSQL-fits, gate-only. This gives PostgreSQL no component-count advantage. MongoDB had 4 kinds and 0 extra processes; PostgreSQL had 12 kinds with the slot feed (9 with outbox) and 6 extra processes. Kinds alone favour MongoDB. The runner applies the gate first, and MongoDB failed F2b's feed gate. Slot was selected most often when labelled F1b lines were included; the eight compared lines split 4–4 between slot and outbox.

The result record includes all 13 compared/labelled lines and their paired numbers. The complete report retains the not-compared variant values, F3's literal outages, gates, feeds, steady cells and component lists. F1b was both-fit at both timings. The labelled floor and outbox-only lines were MongoDB-fits; outbox-only is non-strict, so it does not isolate feed overhead from the strict/non-strict difference.

Strict mode did not settle the durability question

The frozen smoke rule switched PostgreSQL to strict variants after the non-strict smoke observer logged 236 acknowledgements while no synchronous standby was named. Yet the counted strict cells still logged 15,567 acknowledgements in 10 cells while synchronous_standby_names was empty: 3,030 for classic-strict and 12,537 for quorum-strict. Three are quorum-strict F2b cells. The complete report and result record name all ten.

Zero acknowledged writes were lost. The frozen gate counts lost writes, not these observer windows, so verdicts are unchanged. The cause could be a real single-copy interval or the observer reading the wrong node; the cells do not decide it. This run does not establish equal two-copy durability throughout. MongoDB's write gate passed 104/104 cells; its feed passed 101/104. Classic-strict passed 104/104 writes and both compared feeds; quorum-strict passed 116/116. All 14 steady cells passed with no false failovers.

Predictions that the result contradicted

The frozen v1.1 predictions are scored below; the superseded v1 lines remain in the pre-registration. A feed “both-fit” prediction cannot be scored as a verdict because the rule has no separate feed verdict.

Predictions are those frozen with the plan, not a later story about the result.
Who / predictionResult
Design author: the gate passes on both armsMongoDB failed the F2b gate at both timings (feed: 3 of 20 cells, one duplicated effect each); PostgreSQL passed. no.
Design author: F1 / F2a / F2b mongodb-fits, default and tunedF1, F2a mongodb-fits at both; F2b postgres-fits, gate-only, at both. F1, F2a yes; F2b no.
Design author: F3 both-fitmongodb-fits at both timings. no.
Design author: feed both-fit (outbox)the rule has no separate feed verdict; PostgreSQL's outbox and slot passed every cell; MongoDB's stream 101 of 104. not scorable as stated.
Design author: components mongodb-fitspostgres-fits, gate-only (kinds 4 against 12 / 9 meet the kinds test for MongoDB). no by verdict; yes by kinds.
MongoDB advocate: F1 default mongodb-fits, high (about 11–16 s against 30–42 s)mongodb-fits; MongoDB 10.62–20.87 s, PostgreSQL (classic-strict) 25.55–36.59 s. verdict yes; ranges partly.
MongoDB advocate: F1 tuned mongodb-fits, moderate-highmongodb-fits. yes.
MongoDB advocate: F2a / F2b as F1F2a mongodb-fits at both; F2b postgres-fits, gate-only, at both. F2a yes; F2b no.
MongoDB advocate: F3 both-fit, leaning MongoDB, lowmongodb-fits at both timings. no (the lean was right).
MongoDB advocate: components mongodb-fits, highpostgres-fits, gate-only. no by verdict; yes by kinds.
PostgreSQL advocate: F1 / F2a / F2b mongodb-fits, high (tuned about 5.5–7 s against 19–23 s)F1, F2a yes; F2b postgres-fits, gate-only; tuned F1 MongoDB 10.10–10.40 s, PostgreSQL (quorum-strict) 20.52–23.49 s. F1, F2a yes; F2b no; the PostgreSQL range yes, the MongoDB range no.
PostgreSQL advocate: F3 mongodb-fits about 55 % / both-fit about 45 %, lowmongodb-fits at both timings. yes (its 55 % side).
PostgreSQL advocate: feed both-fitas for this PC. not scorable as stated.
PostgreSQL advocate: components mongodb-fits, highpostgres-fits, gate-only. no by verdict; yes by kinds.

Nobody predicted the MongoDB F2b feed failure. Components predictions were wrong by the gate-first verdict, while their expectation of fewer MongoDB kinds was right.

Discards and the unresolved overlap

48 counted-cell attempts were discarded, retained and rerun: 7 MongoDB and 41 PostgreSQL. The complete report reconciles all 48: 2 foreign-work events seen inside a cell; 16 overlaps from other sessions' reports (A/B/C); 4 manually found model loads (D/E/F/G); 6 own-agent test-suite overlaps (H); 20 model-load guard moves (H/K/L/M/N). Earlier drafts counted only 22. The reconciliation includes all seven MongoDB discards, including a different F2b cell from the three feed failures.

Window I remains open: test-suite activity began 8 November at 22:45 local time, with no recorded end. Five retained seed 2 PostgreSQL files may overlap it: quorum-strict default/F3 rep0 and classic-strict default F1/F1b/F2a/F2b rep0. One, classic-strict default/F1 rep0 (35.1979s), is compared. All five paired differences there are negative; removing that pair leaves the sign unchanged. This sensitivity check does not establish a clean run. None is a MongoDB file or in F2b's compared variant. Reported work windows widened by a minute cannot exclude work outside those reports.

Window J stopped all Python processes while the plan waited between cells; no cell was running and none was discarded. Three marked cells left no file and were rerun as pending. Pauses between cells and memory-gate exits are not counted as disturbances. Amendment 03 records the completed run, snapshot restart counts, reconciliation and unresolved limits; it changes no frozen rule.

Sources, corrections and what remains open

Drafts and scientific records are pinned to private revision d5e74de709cfe0ab0e6ef575cdefd28bdaf6ae97. The final supervisor log is pinned separately to the later revision dc62a35fbb3a534738738d983cc24e3359dc9d6f. This publication recount checked all 334 cell identities against Git blobs, summed gate records and checked 240 verdict pair values. It did not rerun a database or verify every one of the 6,020 original manifest entries.

The original frozen checksum manifest is retained unchanged. It lists the earlier disturbance document, not the later 48-file reconciliation; the supervisor log changed too. Source hashes identify these versions separately, and final document hashes identify the public documents. They establish identity, not scientific correctness.

The MongoDB result review and PostgreSQL result review are original model reviews of earlier drafts with vetoes. The corrected drafts incorporated remedies for misleading component credit, F2b attribution, missing predictions and incomplete discards. Those earlier reviews are not final acceptance of this publication.

One machine, simulated partitions and privately held raw evidence; no overall winner or independent replication. The strict-ack observer anomaly and open heavy-work window remain unresolved. The next scientific work is to establish the anomaly's cause and repeat the affected conditions with a complete disturbance record.