1.5.0

This version updates the biometric engine: it adds support for JPEG 2000 images and reduces response time, improving the service's performance.

Engine improvements from version 1.2.0 to 1.5.0

• Significant improvements in processing speed and robustness, enabling operation with lower-resolution images, cropped images, automatic image rotation, etc.

• The Docker image size has been considerably reduced, going from 1.15GB to 211MB.

• Docker vulnerabilities fixed.

Comparison between version 1.2.0 and 1.5.0

The tests were performed on a 16-core AWS machine, using only 8 cores to ensure that no other process affected the comparison.

The service's Docker container was run on that machine and K6 was used to perform Load (load), Stress (stress) and Spike (spike).

Configuration:

• Environment variable FACEPHI_SELPHID_FACIALEXTRACTOR_NUM_THREADS=4.

• In the file config.json, "number_of_threads": 2.

Summary:

• Version 1.5.0 represents a significant performance improvement over 1.2.0 in all test types.

• The average latency in the load test was reduced by 37%, throughput increased by 53% in the same test, and the p(95) percentiles improved steadily.

• Neither version produced HTTP errors.

Load test table

2 VUs in a loop for 15 minutes.

Metrics	v1.2.0	v1.5.0	Δ
Average latency	813.68 ms	511.41 ms	−37.2 %
Min	566.70 ms	420.47 ms	−25.8 %
Median	808.75 ms	510.31 ms	−36.9 %
p(90)	839.83 ms	547.41 ms	−34.8 %
p(95)	843.65 ms	559.02 ms	−33.7 %
Max	994.11 ms	678.97 ms	−31.7 %
Total iterations	2 079	3 177	+52.8 %
Performance	2.31 iter/s	3.53 iter/s	+52.8 %
HTTP errors	0.00 %	0.00 %	—

Stress test table

Up to 4 VUs ramping over 8 min (4 stages).

Metrics	v1.2.0	v1.5.0	Δ
Average latency	773.35 ms	654.23 ms	−15.4 %
Min	424.49 ms	361.91 ms	−14.7 %
Median	584.56 ms	540.95 ms	−7.5 %
p(90)	1 250 ms	994 ms	−20.5 %
p(95)	1 800 ms	1 020 ms	−43.3 %
Max	1 990 ms	1 110 ms	−44.2 %
Total iterations	1 376	1 607	+16.8 %
Performance	2.86 iter/s	3.34 iter/s	+16.8 %
HTTP errors	0.00 %	0.00 %	—

The most notable improvement under stress occurs at the extremes: p(95) drops 780 ms and the maximum decreases by almost 44%. This indicates that version 1.5.0 handles concurrency spikes better.

Spike test table

Up to 4 VUs in an aggressive ramp over 1 min 40 s (3 stages).

Metrics	v1.2.0	v1.5.0	Δ
Average latency	1 090 ms	913.82 ms	−16.2 %
Min	440.28 ms	362.44 ms	−17.7 %
Median	715.47 ms	854.49 ms	+19.4 %
p(90)	1 890 ms	1 430 ms	−24.3 %
p(95)	1 910 ms	1 490 ms	−22.0 %
Max	1 960 ms	1 610 ms	−17.9 %
Total iterations	305	363	+19.0 %
Performance	3.03 iter/s	3.61 iter/s	+19.1 %
HTTP errors	0.00 %	0.00 %	—

Median latency increases in the spike test for 1.5.0 (715 ms → 854 ms). Although the high percentiles improve, the central distribution worsens slightly below the peak.

Overall comparative summary

Type	1.2.0 average	1.5.0 average	1.2.0 p(95)	1.5.0 p(95)	1.2.0 RPS	1.5.0 RPS
Load	813 ms	511 ms	843 ms	559 ms	2.31	3.53
Stress	773 ms	654 ms	1 800 ms	1 020 ms	2.86	3.34
Spike	1 090 ms	914 ms	1 910 ms	1 490 ms	3.03	3.61