Performance¶

Measured on 2026-10-14 on a 16-core x86_64 Linux host, Python 3.12, NumPy 2, one BLAS thread per process, with the bundled 10 s, 500 Hz, 12-lead examples (python tools/benchmark_throughput.py --records 48 --workers 1 4 8 12). Numbers vary with the host and its load; the CI budget (benchmarks/performance/budget.json) allows about 2.5 times these values.

One record¶

Step	0.1.0	0.2.0	0.3.x (compiled engine)
import ecgfeat	—	0.02 s (lazy; no NumPy until used)	0.01 s
first ecg_record in a process	—	4.0 s	3.6 s
warm ecg_record (median of 5)	2.44 s	2.60 s	2.62 s
query_measurement on a record	16.7 ms	0.03 ms	0.03 ms
ECGRecord.to_rows() (≈4,000 cells)	—	about 10 ms	about 10 ms

The 0.3 column was measured on 2026-10-19 on the same host (load average about 5) with the ecg-records-engine 0.3.0 wheel from PyPI. Compiling the engine with Cython changes nothing measurable: the time is in NumPy and SciPy calls, not in Python bytecode.

Extraction is about 6% slower than 0.1.0 because the engine does more (the measurement corrections since 0.1.0, schema 1.1 fields). Profiling shows no single hotspot: delineation and the atrial analysis take most of the time, spread over many small NumPy calls. Reading records is much faster: queries, tables and plots now share one cached document per record instead of copying the whole record on every call.

Many records¶

ecg-record batch --jobs N (and ecg-record batch DIR) runs N worker processes; 0.1 used threads, which the interpreter lock serialised.

Workers	48 records	Records per hour
1	128 s	1,347
4	34 s	5,067
8	18 s	9,424
12	13 s	13,159

On 24 records with 8 workers, processes took 15.6 s and threads 91.0 s. At 12 workers the 21,799 PTB-XL records take about 1.7 hours.

Interpretation¶

With the separate ecginterpret distribution, ecginterpret.analyze(signal) measures once and returns the record and the Interpretation document: about 2.7 s per record instead of about 5.5–7 s for ecg_record followed by interpret_record.

Tips¶

• Use profile="summary" (default) unless you need the all fields; the JSON is about half the size.
• Install the performance extra (Numba) to JIT the Kalman smoother; the first call in each process pays the compilation.
• For long recordings, ecg_record_segments measures 10 s windows; run windows in parallel with your own process pool if needed.