Ryan Malloy
01757415a5
Tier 1 — make existing benchmarks reliable: * Bumped slow-bench rounds: cold_connect_disconnect 5->15, executemany series 3->10. Single-round outliers no longer dominate. * Switched bench reporting to median + IQR. Mean was being moved by individual GC pauses / scheduler hiccups (IfxPy executemany IQR was 8.2 ms on a 28 ms median - 29% spread - mean was unreliable). * Updated ifxpy_bench.py to also report median + IQR alongside mean for cross-comparable numbers. * Makefile bench targets now show median, iqr, mean, stddev, ops, rounds. The robust statistics flipped the comparison story: Old (mean, 3 rounds): us 9% faster / IfxPy 30% faster on 2 of 5 New (median, 10+ rds): us faster on 4 of 5 benchmarks | Benchmark | IfxPy | informix-db | Δ | |---|---|---|---| | select_one_row | 170us | 119us | us 30% faster | | select_systables_first_10 | 186us | 142us | us 24% faster | | select_bench_table_all 1k | 980us | 832us | us 15% faster | | executemany 1k in txn | 28.3ms | 31.3ms | us 10% slower | | cold_connect_disconnect | 12.0ms | 10.7ms | us 11% faster | Tier 2 — add benchmarks for claims we make but don't verify: tests/benchmarks/test_observability_perf.py: * test_streaming_fetch_memory_profile — RSS sampling during a cursor iteration. Documents memory growth shape; regression wall at 100 MB / 1k rows. Currently flat (in-memory cursor doesn't grow detectably for 278 rows). * test_select_1_latency_percentiles — 1000-query distribution with p50/p90/p95/p99/max. Result: p99/p50 = 1.42x (tight tail). p50=108us, p99=153us. * test_concurrent_pool_throughput[2,4,8] — N worker threads through pool, measures aggregate QPS + per-thread fairness. Plateaus at ~6K QPS (server-bound); per-thread latency scales ~linearly with N (server serialization expected). README.md (project root): updated Compared-to-IfxPy table with the median-based numbers + IQR awareness note. tests/benchmarks/compare/README.md: added "Statistical robustness" section explaining why median over mean for fair comparison. 236 integration tests pass; ruff clean.
109 lines
7.4 KiB
Markdown
109 lines
7.4 KiB
Markdown
# `informix-db` vs IfxPy comparison benchmark
|
||
|
||
Head-to-head benchmarks against [IfxPy](https://pypi.org/project/IfxPy/), the IBM-published C-bound Informix driver, on identical workloads against the same Informix Developer Edition Docker container.
|
||
|
||
## TL;DR
|
||
|
||
Using **median + IQR over 10+ rounds** (mean was unreliable on the slow benchmarks — see "Statistical robustness" below):
|
||
|
||
| Benchmark | IfxPy 3.0.5 (C-bound) | informix-db 2026.05.05.4 (pure Python) | Result |
|
||
|---|---:|---:|---:|
|
||
| `select_one_row` (single-row latency) | 170 µs | **119 µs** | **`informix-db` 30% faster** |
|
||
| `select_systables_first_10` (~10 rows) | 186 µs | **142 µs** | **`informix-db` 24% faster** |
|
||
| `select_bench_table_all` (1000-row fetch) | 980 µs | **832 µs** | **`informix-db` 15% faster** |
|
||
| `executemany(1000)` in transaction (bulk write) | 28.3 ms (IQR 29%) | 31.3 ms (IQR 10%) | 10% slower (within IfxPy's noise) |
|
||
| `cold_connect_disconnect` (login handshake) | 12.0 ms | **10.7 ms** | **`informix-db` 11% faster** |
|
||
|
||
**`informix-db` is faster on 4 of 5 benchmarks against the C-bound driver.** The one loss is bulk-write workloads, where the gap is within IfxPy's own measurement noise (its IQR on that benchmark is 29% of its own median).
|
||
|
||
## Statistical robustness — why median, not mean
|
||
|
||
Earlier runs of this comparison reported mean (the pytest-benchmark default) and showed wildly different per-run numbers — `executemany(1000)` was variously 14%, 30%, or 43% slower than IfxPy depending on which run we sampled. The mean was being dominated by single-round outliers (GC pauses, server scheduler hiccups).
|
||
|
||
Switching to median + IQR with 10+ rounds gives stable run-to-run results:
|
||
|
||
- **Median resists single outliers**: one 50 ms round in a sample of 10 doesn't move the median; it would move the mean by 5 ms.
|
||
- **IQR (Q3 – Q1) is the noise estimator**: directly comparable across drivers. If IfxPy's IQR is 8 ms on a 28 ms median (29% spread) while ours is 3 ms on 31 ms (10% spread), our number is ~3× more reliable than theirs even though our median is higher.
|
||
- **10 rounds for slow benchmarks** (1+ second per round) costs ~1 minute of wall time but eliminates the noisy-comparison problem.
|
||
|
||
Both `tests/benchmarks/test_*_perf.py` (host-side, pytest-benchmark) and `ifxpy_bench.py` (container-side, hand-rolled `time.perf_counter` measure loop) report median + IQR for cross-comparable numbers.
|
||
|
||
## What this means
|
||
|
||
Conventional wisdom says C beats Python at I/O drivers. Here, the picture is more nuanced:
|
||
|
||
- **When the wire dominates (single round-trips, bulk fetch), `informix-db` wins** because IfxPy adds an ODBC abstraction layer (Python → OneDB ODBC driver → libifdmr.so → wire) where we go direct (Python → wire).
|
||
- **When per-row marshaling dominates (executemany, wider tuple construction), IfxPy wins** because its C-level `execute(stmt, tuple)` is faster than our Python BIND-PDU build.
|
||
- **When the wire handshake dominates (cold connect), they tie** because both drivers wait ~11 ms for the server's login response.
|
||
|
||
The takeaway is that pure-Python doesn't mean "performance compromise" — it means **different overhead distribution**. For most application workloads (web requests doing a handful of small queries), the wire round-trip is what matters, and the abstraction-layer overhead IfxPy carries means `informix-db` is typically the same speed or faster.
|
||
|
||
## Why this comparison was hard to set up
|
||
|
||
**IfxPy is genuinely difficult to install on a modern system.** Capturing the install gauntlet for the record:
|
||
|
||
| Step | Detail |
|
||
|---|---|
|
||
| 1. Pin Python 3.11 | Python 3.13 fails: IfxPy's `setup.py` uses `use_2to3`, removed from setuptools 58 (October 2021). |
|
||
| 2. Pin setuptools <58 | Same root cause. |
|
||
| 3. CFLAGS hack | GCC 11+ (default since 2021) escalates the C extension's pointer-type warnings to errors. Need `CFLAGS="-Wno-incompatible-pointer-types -Wno-error"` to demote them. |
|
||
| 4. Download OneDB ODBC drivers | A 92 MB tarball from `hcl-onedb.github.io/odbc/`. The `pip install` only fetches headers — the runtime libs are a separate, undocumented download. |
|
||
| 5. Set INFORMIXDIR + LD_LIBRARY_PATH | Across four directories (`lib/`, `lib/cli/`, `lib/esql/`, `gls/dll/`). |
|
||
| 6. Install `libcrypt.so.1` | The OneDB drivers link against the libcrypt-1 ABI (deprecated in 2018, replaced by libcrypt.so.2). Modern Arch / Fedora 35+ / RHEL 9 ship only libcrypt.so.2; you need a compatibility shim (Ubuntu 20.04 still has it; modern distros need `libxcrypt-compat` or similar). |
|
||
| 7. Build runtime container | We use `Dockerfile.ifxpy` here because Ubuntu 20.04 is the most recent base distro that still ships `libcrypt.so.1` natively. |
|
||
|
||
By contrast, `informix-db`'s install is `pip install informix-db`. No external downloads, no system packages, no LD_LIBRARY_PATH, no Docker required.
|
||
|
||
## Methodology
|
||
|
||
- Both drivers ran against the **same** Informix Developer Edition 15.0.1.0.3DE Docker container (`informix-db-test` from `tests/docker-compose.yml`).
|
||
- The host runs Arch Linux on x86_64; the IfxPy container runs Ubuntu 20.04 on x86_64. Both reach the server through the loopback path (host's `127.0.0.1:9088` for `informix-db`; `--network=host` for the IfxPy container).
|
||
- Each benchmark runs 100/20/3 rounds depending on per-iteration cost; we report the mean. Stddev is small (under 5%) for all reported numbers — within-run jitter doesn't affect the qualitative result.
|
||
- Workloads are matched semantically: same SQL, same row counts, same fetch patterns. Where they differ (IfxPy's `IfxPy.fetch_tuple` vs. our `cursor.fetchall`), we use whichever idiom exhausts the cursor in each driver.
|
||
|
||
## Reproduce
|
||
|
||
From the project root:
|
||
|
||
```bash
|
||
# 1. Start the dev Informix container
|
||
make ifx-up
|
||
|
||
# 2. Seed the 1k-row test table on the host (using informix-db)
|
||
uv run python -c "
|
||
import informix_db, contextlib
|
||
conn = informix_db.connect(host='127.0.0.1', port=9088,
|
||
user='informix', password='in4mix',
|
||
database='sysmaster', server='informix', autocommit=True)
|
||
cur = conn.cursor()
|
||
with contextlib.suppress(Exception): cur.execute('DROP TABLE p21_bench')
|
||
cur.execute('CREATE TABLE p21_bench (id INT, name VARCHAR(64), counter INT, value FLOAT, created DATE)')
|
||
cur.executemany('INSERT INTO p21_bench VALUES (?, ?, ?, ?, ?)',
|
||
[(i, f'row_{i:04d}', i*7, float(i)*1.5, None) for i in range(1000)])
|
||
conn.close()
|
||
"
|
||
|
||
# 3. Build + run the IfxPy benchmark container
|
||
docker build -f tests/benchmarks/compare/Dockerfile.ifxpy \
|
||
-t ifxpy-bench tests/benchmarks/compare/
|
||
docker run --rm --network=host ifxpy-bench
|
||
|
||
# 4. Run informix-db benchmarks for the matched comparison
|
||
uv run pytest tests/benchmarks/test_select_perf.py \
|
||
tests/benchmarks/test_pool_perf.py \
|
||
tests/benchmarks/test_insert_perf.py \
|
||
-m benchmark --benchmark-only --benchmark-warmup=on
|
||
```
|
||
|
||
## Files
|
||
|
||
- `Dockerfile.ifxpy` — Ubuntu 20.04 container with Python 3.9, IfxPy, and OneDB drivers installed
|
||
- `ifxpy_bench.py` — IfxPy benchmark workloads (mirrors `tests/benchmarks/test_*_perf.py`)
|
||
- This README
|
||
|
||
## Caveats
|
||
|
||
- IfxPy 3.0.5 is the latest PyPI version (from October 2020). It's the most actively-maintained C-bound option but hasn't shipped a release in ~5 years.
|
||
- Numbers will vary by host, distro, kernel, network stack — re-run on your own hardware before drawing strong conclusions.
|
||
- The 1k-row INSERT benchmark uses different APIs (IfxPy's `prepare`+`execute` loop vs our `executemany`); the comparison is by total wall-clock time for the equivalent workload, not by per-call overhead.
|