Sandbox File System I/O Benchmarks

Benchmarks SQLite performance across five sandbox providers: Tensorlake, Vercel, Daytona, E2B, and Modal.

All sandboxes were configured with 2 vCPUs and ~4 GB RAM. Each benchmark was run 3 times, reporting mean +/- standard deviation.

What It Benchmarks

The benchmark script (benchmark.py) runs 11 SQLite operations with deterministic data (random.seed(42)) across three modes:

Mode	Description
Default	WAL mode, synchronous=NORMAL, 64MB cache, 60k rows (8.68 MB)
Fsync	synchronous=FULL, DELETE journal mode, 25k rows (4.05 MB)
Large	WAL mode, synchronous=NORMAL, 8MB cache, 250k rows (34.85 MB)

Operation	Description
Sequential inserts	Single-row INSERTs in autocommit
Batch inserts	executemany INSERT
SELECT COUNT(*)	Full table count
Range queries	WHERE value BETWEEN ? AND ? with index
LIKE queries	WHERE name LIKE ? (full scan)
Updates	Single-row UPDATEs by name
Deletes	Single-row DELETEs by name
Transaction inserts	INSERTs within explicit BEGIN/COMMIT
Aggregates	AVG, MIN, MAX, SUM + GROUP BY
Join query	Two-table JOIN with WHERE filter
Concurrent reads	4 threads x 500 range queries each (tests multi-core)

Results

Default Mode (WAL, 60k rows, 8.68 MB)

Provider	Total Time	Concurrent Reads	Reads q/s
Tensorlake	2.45s +/- 0.04s	1.17s	1711 q/s
Vercel	3.00s +/- 0.09s	1.37s	1465 q/s
E2B	3.92s +/- 0.28s	1.73s	1175 q/s
Modal	4.66s +/- 0.19s	2.31s	869 q/s
Daytona	5.51s +/- 0.18s	3.90s	511 q/s

Tensorlake  ██████████████████████████████  2.45s
Vercel      ████████████████████████        3.00s
E2B         █████████████████████           3.92s
Modal       ███████████████████             4.66s
Daytona     █████████████                   5.51s

Detailed breakdown (seconds)

Benchmark	Tensorlake	Vercel	Daytona	E2B	Modal
Sequential inserts	0.0498	0.0635	0.0577	0.0808	0.0913
Batch inserts	0.2514	0.3115	0.2957	0.3868	0.4756
SELECT COUNT(*)	0.0001	0.0001	0.0001	0.0001	0.0001
Range queries	0.1018	0.1286	0.0946	0.1615	0.1473
LIKE queries	0.8154	1.0524	1.0858	1.4749	1.5034
Updates	0.0123	0.0153	0.0147	0.0184	0.0263
Deletes	0.0050	0.0060	0.0068	0.0071	0.0157
Transaction inserts	0.0246	0.0337	0.0385	0.0381	0.0577
Aggregates	0.0147	0.0207	0.0159	0.0222	0.0229
Join query	0.0014	0.0021	0.0017	0.0027	0.0021
Concurrent reads	1.1711	1.3680	3.8962	1.7317	2.3127

Fsync Mode (synchronous=FULL, DELETE journal, 25k rows, 4.05 MB)

Provider	Total Time	Concurrent Reads	Reads q/s
Tensorlake	1.64s +/- 0.20s	1.31s	1558 q/s
Vercel	2.05s +/- 0.11s	1.61s	1242 q/s
E2B	2.17s +/- 0.66s	1.58s	1175 q/s
Modal	2.43s +/- 0.11s	1.79s	1120 q/s
Daytona	4.75s +/- 0.90s	4.33s	475 q/s

Tensorlake  ██████████████████████████████  1.64s
Vercel      ██████████████████████████      2.05s
E2B         █████████████████████████       2.17s
Modal       ████████████████████████        2.43s
Daytona     ██████████                      4.75s

Detailed breakdown (seconds)

Benchmark	Tensorlake	Vercel	Daytona	E2B	Modal
Sequential inserts	0.0244	0.0323	0.0407	0.0410	0.0465
Batch inserts	0.0901	0.1180	0.1118	0.1502	0.1709
SELECT COUNT(*)	0.0000	0.0001	0.0001	0.0001	0.0002
Range queries	0.0472	0.0604	0.0451	0.0896	0.0832
LIKE queries	0.1353	0.1739	0.1757	0.2437	0.2543
Updates	0.0047	0.0080	0.0089	0.0075	0.0138
Deletes	0.0023	0.0055	0.0075	0.0044	0.0116
Transaction inserts	0.0215	0.0294	0.0272	0.0390	0.0435
Aggregates	0.0065	0.0092	0.0069	0.0101	0.0102
Join query	0.0014	0.0019	0.0017	0.0025	0.0026
Concurrent reads	1.3054	1.6145	4.3252	1.5774	1.7909

Large Mode (WAL, 250k rows, 34.85 MB, 8MB cache)

Provider	Total Time	Concurrent Reads	Reads q/s
Tensorlake	9.45s +/- 0.10s	1.20s	1681 q/s
Vercel	11.97s +/- 0.05s	1.20s	1666 q/s
E2B	15.69s +/- 0.63s	1.63s	1273 q/s
Daytona	16.69s +/- 0.88s	4.05s	509 q/s
Modal	63.55s +/- 0.36s	19.72s	101 q/s

Tensorlake  ██████████████████████████████  9.45s
Vercel      ███████████████████████         11.97s
E2B         ██████████████████              15.69s
Daytona     █████████████████               16.69s
Modal       ████                            63.55s

Detailed breakdown (seconds)

Benchmark	Tensorlake	Vercel	Daytona	E2B	Modal
Sequential inserts	0.2608	0.3523	0.3175	0.4267	0.6368
Batch inserts	1.0235	1.2737	1.2111	1.6318	2.8765
SELECT COUNT(*)	0.0001	0.0004	0.0003	0.0004	0.0007
Range queries	0.3102	0.4084	0.4263	0.4842	1.5284
LIKE queries	6.4741	8.4868	10.4028	11.2522	38.0541
Updates	0.0368	0.0466	0.0579	0.0458	0.1492
Deletes	0.0243	0.0403	0.0657	0.0310	0.1553
Transaction inserts	0.0491	0.0687	0.0733	0.0829	0.1661
Aggregates	0.0747	0.0835	0.0777	0.1036	0.2524
Join query	0.0020	0.0033	0.0032	0.0038	0.0107
Concurrent reads	1.1965	1.2016	4.0498	1.6261	19.7180

Key Observations

• Tensorlake wins all three modes, consistently fastest across every dataset size and configuration.
• Vercel is a solid second place overall, with tight competition from E2B in fsync mode.
• Daytona concurrent reads are 3-4x slower than every other provider across all modes. This is the single biggest performance differentiator -- Daytona's single-threaded operations are competitive, but multi-threaded reads expose a bottleneck.
• Modal's large mode performance collapsed -- 6.7x slower than Tensorlake (63.55s vs 9.45s). LIKE queries took 38s (vs 6.5s on Tensorlake) and concurrent reads hit only 101 q/s. Modal uses SQLite 3.40.1 (oldest in the set), which may explain the degradation at scale. Default and fsync modes were mid-pack.
• LIKE queries (full table scans) dominate total time in all modes, especially at the large scale where they account for 60-70% of runtime.

Environment

	Tensorlake	Vercel	Daytona	E2B	Modal
Python	3.12.3	3.13.1	3.13.12	3.13.12	3.13.3
SQLite	3.45.1	3.51.1	3.46.1	3.46.1	3.40.1
vCPUs (verified)	2	2	2 (cgroup)	2	2
Memory (verified)	3.9 GB	4.3 GB	4.0 GB (cgroup)	3.9 GB	~1 TB (host)

Running the Benchmarks

Prerequisites

Install and authenticate each provider's CLI:

Provider	Install	Auth
Tensorlake	pip install tensorlake (into /tmp/venv)	tensorlake login
Vercel	npm i -g sandbox	sandbox login
Daytona	brew install daytonaio/cli/daytona	daytona login
E2B	npm i -g e2b	e2b auth login
Modal	pip install modal	modal token set

E2B requires building a template to configure CPU and memory:

mkdir /tmp/e2b-template
echo 'FROM python:3.13-slim' > /tmp/e2b-template/Dockerfile
cd /tmp/e2b-template
e2b template create bench-2cpu-4gb \
  --dockerfile Dockerfile \
  --cpu-count 2 \
  --memory-mb 4096

Usage

# Run all providers, all modes (default + fsync + large)
python run_benchmarks.py

# Run specific providers
python run_benchmarks.py tensorlake vercel

# Use a custom E2B template
python run_benchmarks.py e2b --e2b-template bench-2cpu-4gb

Provider Notes

• Vercel: Python runtime does not include the native _sqlite3 C extension. The runner installs pysqlite3-binary automatically.
• Daytona: Using --class small locks you to a snapshot with fixed 1 vCPU. To get custom resources, use --cpu/--memory with -f Dockerfile instead. nproc reports host CPUs, but cgroup enforces the requested limit.
• E2B: CPU and memory cannot be set at sandbox creation time. You must build a custom template with e2b template create --cpu-count N --memory-mb N.
• Modal: Python SDK (not CLI). Sandboxes are created via modal.Sandbox.create(). Requires pip install modal and modal token set. Memory reports ~1 TB as it sees the host; actual sandbox allocation may differ.

Files

benchmark.py          # SQLite benchmark (runs inside sandboxes)
run_benchmarks.py     # Orchestrator: create, copy, run, collect, cleanup
results/
  results.json        # Benchmark results (2 vCPU / 4 GB, March 17 2026)