# Backend Comparison 2026-03-17

## Purpose

This document compares the current Oracle, PostgreSQL, and MongoDB workflow-engine backends using the published normalized performance baselines and the currently implemented backend-specific validation suites.

This is a decision pack for the current local Docker benchmark set. It is not the final production recommendation pack yet, because hostile-condition and Bulstrad hardening depth is still strongest on Oracle.

The durable machine-readable companion is [13-backend-comparison-2026-03-17.json](13-backend-comparison-2026-03-17.json).

For the exact six-profile signal-driver matrix, including `Oracle+Redis`, `PostgreSQL+Redis`, and `Mongo+Redis`, see [14-signal-driver-backend-matrix-2026-03-17.md](14-signal-driver-backend-matrix-2026-03-17.md).

## Source Baselines

- [10-oracle-performance-baseline-2026-03-17.md](10-oracle-performance-baseline-2026-03-17.md)
- [11-postgres-performance-baseline-2026-03-17.md](11-postgres-performance-baseline-2026-03-17.md)
- [12-mongo-performance-baseline-2026-03-17.md](12-mongo-performance-baseline-2026-03-17.md)

## Validation Status

Current comparison-relevant validation state:

- Oracle performance suite: `12/12` passed
- PostgreSQL performance suite: `11/11` passed
- MongoDB performance suite: `14/14` passed
- PostgreSQL focused backend parity suite: `9/9` passed
- MongoDB focused backend parity suite: `23/23` passed
- integration project build: `0` warnings, `0` errors

Important caveat:

- Oracle still has the broadest hostile-condition and Bulstrad E2E matrix.
- PostgreSQL and MongoDB now have backend-native signal/runtime/projection/performance coverage plus curated Bulstrad parity, but they do not yet match Oracle's full reliability surface.

## Normalized Comparison

### Synthetic Signal Round-Trip

| Metric | Oracle | PostgreSQL | MongoDB |
| --- | ---: | ---: | ---: |
| Serial latency avg ms | 3104.85 | 3079.33 | 97.88 |
| Serial latency P95 ms | 3165.04 | 3094.94 | 149.20 |
| Throughput ops/s | 20.98 | 25.74 | 76.28 |
| Throughput avg ms | 4142.13 | 3603.54 | 1110.94 |
| Throughput P95 ms | 4215.64 | 3635.59 | 1121.22 |
| Soak ops/s | 3.91 | 4.30 | 47.62 |
| Soak avg ms | 4494.29 | 4164.52 | 322.40 |
| Soak P95 ms | 5589.33 | 4208.42 | 550.50 |

### Capacity Ladder

| Concurrency | Oracle ops/s | PostgreSQL ops/s | MongoDB ops/s |
| --- | ---: | ---: | ---: |
| `c1` | 3.37 | 4.11 | 7.08 |
| `c4` | 15.22 | 17.29 | 38.35 |
| `c8` | 21.34 | 33.21 | 66.04 |
| `c16` | 34.03 | 57.04 | 68.65 |

### Transport Baselines

| Scenario | Oracle ops/s | PostgreSQL ops/s | MongoDB ops/s |
| --- | ---: | ---: | ---: |
| Immediate burst nightly | 50.18 | 70.10 | 46.18 |
| Delayed burst nightly | 10.71 | 19.60 | 16.66 |
| Immediate burst smoke | 56.94 | 70.21 | 32.11 |
| Delayed burst smoke | 2.86 | 5.59 | 4.97 |

### Bulstrad Workloads

| Scenario | Oracle ops/s | PostgreSQL ops/s | MongoDB ops/s |
| --- | ---: | ---: | ---: |
| `QuoteOrAplCancel` smoke | 19.69 | 59.88 | 23.48 |
| `QuotationConfirm -> ConvertToPolicy` nightly | 1.77 | 3.33 | 10.83 |

## Backend-Specific Observations

### Oracle

- Strongest validation depth and strongest correctness story.
- Main cost center is still commit pressure.
- Dominant wait is `log file sync` in almost every measured scenario.
- `c8` is still the last comfortable rung on the local Oracle Free setup.

### PostgreSQL

- Best relational performance profile in the current measurements.
- Immediate transport is the strongest of the three measured backends.
- Dominant wait is `Client:ClientRead`, which points to queue-claim cadence and short-transaction wake behavior, not a clear storage stall.
- `c16` still scales cleanly on this benchmark set and does not yet show a hard saturation cliff.

### MongoDB

- Fastest measured backend across the synthetic signal round-trip workloads and the medium Bulstrad nightly flow.
- No stable top wait dominated the measured runs; current analysis is more meaningful through normalized metrics and Mongo counters than through wait classification.
- The significant findings were correctness issues discovered and fixed during perf work:
  - bounded empty-queue receive
  - explicit collection bootstrap before transactional concurrency
- `c16` is the first rung where latency visibly expands even though throughput still rises.

## Decision View

### Raw Performance Ranking

For the current local benchmark set:

1. MongoDB
2. PostgreSQL
3. Oracle

This order is stable for:

- serial latency
- steady-state synthetic throughput
- soak throughput
- medium Bulstrad nightly flow

### Validation Maturity Ranking

For the current implementation state:

1. Oracle
2. PostgreSQL
3. MongoDB

Reason:

- Oracle has the deepest hostile-condition and Bulstrad E2E surface.
- PostgreSQL now has a solid backend-native suite and competitive performance, but less reliability breadth than Oracle.
- MongoDB now has good backend-native and performance coverage, but its operational model is still the most infrastructure-sensitive because it depends on replica-set transactions plus change-stream wake behavior.

### Current Recommendation

If the next backend after Oracle must be chosen today:

- choose PostgreSQL as the next default portability target

Reason:

- it materially outperforms Oracle on the normalized workflow workloads
- it preserves the relational operational model for runtime state and projections
- its wake-hint model is simpler to reason about operationally than MongoDB change streams
- it now has enough backend-native correctness and Bulstrad parity coverage to be a credible second backend

If the decision is based only on benchmark speed:

- MongoDB is currently fastest

But that is not the same as the safest operational recommendation yet.

## What Remains Before A Final Production Recommendation

- expand PostgreSQL hostile-condition coverage to match the broader Oracle matrix
- expand MongoDB hostile-condition coverage to match the broader Oracle matrix
- widen the curated Bulstrad parity pack further on PostgreSQL and MongoDB
- rerun the shared parity pack on all three backends in one closeout pass
- add environment-to-environment reruns before turning local Docker numbers into sizing guidance

## Short Conclusion

The engine is now backend-comparable on normalized performance across Oracle, PostgreSQL, and MongoDB.

The current picture is:

- Oracle is the most validated backend
- PostgreSQL is the best performance-to-operability compromise
- MongoDB is the fastest measured backend but not yet the safest backend recommendation