MongoDB Transaction Performance Benchmark

Background

While discussing database performance, an LLM told me that MongoDB with transactions runs 5-15x slower than without transactions. This seemed like a substantial claim worth investigating, so I decided to test it.

I vibe coded this benchmark to compare MongoDB's performance with and without transactions on a single machine. All replica set members run on localhost, which means essentially zero network latency between them. The tests revealed a ~2x slowdown for typical workloads, but this overhead increases dramatically with larger transaction sizes, reaching ~7x at 40,000 operations per transaction. It's important to note that production setups with replica set members on different physical machines would experience additional overhead from network latency.

After running my tests and, I found this EnterpriseDB report which appears to be the source of the 5-15x figure. The report actually compares MongoDB transactions versus PostgreSQL transactions, not MongoDB with versus without transactions. This explained the discrepancy.

So TL;DR is that mongo is indeed kinda slow, but the LLM misrepresented the number.

Results

Testing on a machine with the total of 12 cores, MongoDB 8.0:

Test 1: 1M prepopulated, 500K inserts (10 workers, baseline)

Mode	Throughput	Avg Latency	Duration
Without TX	457,961 ops/s	20.6 ms	1.09s
With TX	228,646 ops/s	43.1 ms	2.19s

Slowdown: ~2.0x

Test 2: 10M prepopulated, 10M inserts (10 workers, 2 indexes, 10 batches/tx)

Mode	Throughput	Avg Latency/TX	Duration
Without TX	239,573 ops/s	416.4 ms	41.74s
With TX	92,579 ops/s	1077.8 ms	1m48.02s

Slowdown: ~2.6x

Test 3: 5M prepopulated, 5M mixed ops (10 workers, 3 indexes, 30 batches/tx, cross-collection)

Mode	Throughput	Avg Latency/TX	Duration
Without TX	197,708 ops/s	1495.1 ms	25.29s
With TX	73,016 ops/s	4061.7 ms	1m8.48s

Slowdown: ~2.7x

Test 3b: 5M prepopulated, 5M mixed ops (10 workers, 3 indexes, 40 batches/tx, cross-collection)

Mode	Throughput	Avg Latency/TX	Duration
Without TX	196,001 ops/s	1992.5 ms	25.51s
With TX	27,562 ops/s	14401.1 ms	3m1.41s

Slowdown: ~7.1x - Transaction overhead increases dramatically with larger transaction sizes (40K ops/tx vs 30K ops/tx).

Usage

Prerequisites: Go 1.23+, Docker or Podman engine running

# Run both tests (default: fair comparison)
go test -v -run TestBatchInsert

# Run specific test
go test -v -run TestBatchInsert_NoTx
go test -v -run TestBatchInsert_Tx

Configuration

Basic Parameters

Variable	Default	Description
PREPOPULATE_DOCS	1000000	Documents to prepopulate
INSERT_BATCHES	500	Number of batch operations
BATCH_SIZE	1000	Documents per batch
INSERT_WORKERS	10	Concurrent workers

Workload Parameters

Variable	Default	Description
MIXED_WORKLOAD	0	Enable mixed read+update+insert workload (set to 1) When enabled: 25% reads, 25% updates, 50% inserts

Stress Parameters (to increase TX overhead)

Variable	Default	Description
SECONDARY_INDEXES	0	Number of secondary indexes (try 2-4)
BATCHES_PER_TX	1	Batches per transaction (try 10-100)
CROSS_COLLECTION	0	Enable cross-collection writes (set to 1)

Progress

Variable	Default	Description
PREPOP_PROGRESS_SECONDS	2	Prepopulation progress interval
TEST_PROGRESS_SECONDS	1	Insert progress interval

Example Scenarios

Baseline (insert-only, minimal overhead)

go test -v -run TestBatchInsert

Heavy Load

PREPOPULATE_DOCS=10000000 INSERT_BATCHES=10000 \
SECONDARY_INDEXES=2 BATCHES_PER_TX=10 \
go test -v -run TestBatchInsert

Stress Test (mixed workload + indexes + 30K ops/tx)

PREPOPULATE_DOCS=5000000 INSERT_BATCHES=5000 MIXED_WORKLOAD=1 SECONDARY_INDEXES=3 BATCHES_PER_TX=30 CROSS_COLLECTION=1 \
go test -v -run TestBatchInsert

Extreme Stress Test (40K ops/tx - shows non-linear overhead)

PREPOPULATE_DOCS=5000000 INSERT_BATCHES=5000 MIXED_WORKLOAD=1 SECONDARY_INDEXES=3 BATCHES_PER_TX=40 CROSS_COLLECTION=1 \
go test -v -run TestBatchInsert