Repository navigation
Releases
QFastServer 3.44.15-qfs.2
Second release of QFastServer for QGIS Server 3.44.15: empty maps are written as compact PNG files, and the
image can start several QGIS Server processes.
docker pull ghcr.io/envirosolutionspl/qfastserver:3.44.15-qfs.2Use the image exactly as qgis/qgis-server:3.44.15-noble: same ports, volumes and environment variables.
Without the two new options the container is started by the unchanged start script of the official image.
Compatibility
| QGIS Server | 3.44.15 |
| Upstream tag and commit | final-3_44_15, 44b4ef051784192d6e4829dc008bca120e4d2388 |
| Patch series | 3.44, 15 patches, fork commit 8da1b23a4bdc992d19575be2097af3c4f3a16c81 |
| Base image | qgis/qgis-server:3.44.15-noble |
| Qt | 5.15 |
Verification
- Responses compared with the official image
qgis/qgis-server:3.44.15-noble, 59 request types, both images
started with their defaults (one process), four passes: 55 of 59 identical in each pass. The four others
(GetContext, 8 bits PNG, DXF, PDF) also differ between two processes of the official image itself, which
compared with itself gave 56 of 59. - PNG maps compared with the release 3.44.15-qfs.1, 10 kinds of maps: the decoded pixels are the same for all
of them; the files of the maps which are not nearly empty are the same byte for byte. - Upstream tests of the server: 56 of 56, the three which need PostgreSQL included. They were run with the
same patches on the branchrelease-3_4428 commits after the tagfinal-3_44_15, not on the tag itself.
The suites for rasters, projects, vector layers and DXF were not run again: the new patch changes the PNG
writer of the WMS service only. The 68 raster renderings were not compared again for the same reason. - Start of the image: with the defaults, the same processes as the official image. With
QFASTSERVER_WORKERSthe number of processes asked for, also together withSKIP_NGINX,
NGINX_CORS_ALLOW_ORIGINandQGIS_SERVER_LANDING_PAGE_PREFIXof the official image; the container stops
when one of the processes ends. WithQFASTSERVER_GZIPa GetCapabilities document is the same after
decompression.
Performance
Release 3.44.15-qfs.1 against this release, both images with their defaults in a container with 4 CPUs and
4 GB, PostGIS 16 on the same host. One client, the two images alternating, median of 30 requests.
| Map, PNG | Size, qfs.1 → qfs.2 | Median, qfs.1 → qfs.2 |
|---|---|---|
| Empty, transparent, 1920x1080 | 36335 → 8151 bytes | 33.5 → 33.9 ms |
| Empty, white background, 1920x1080 | 38677 → 9995 bytes | 32.7 → 33.7 ms |
| Empty tile, transparent, 256x256 | 1246 → 355 bytes | 4.4 → 4.6 ms |
| Points with labels, transparent, 1920x1080 | 37516 → 9465 bytes | 33.5 → 33.6 ms |
| A few roads, transparent, 1920x1080 | 46856 → 16728 bytes | 33.3 → 33.9 ms |
| Roads, transparent, 1920x1080 | 168533 → 168533 bytes | 40.7 → 40.7 ms |
| Buildings, transparent, 1920x1080 | 323334 → 323334 bytes | 47.7 → 48.0 ms |
| 3 vector layers, white background, 1920x1080 | 546993 → 546993 bytes | 60.3 → 61.4 ms |
| 3 vector layers, white background, 1024x768 | 304906 → 304906 bytes | 29.9 → 29.9 ms |
| Tile with buildings, transparent, 256x256 | 23196 → 23196 bytes | 5.7 → 5.8 ms |
The official image writes the empty transparent map of 1920x1080 pixels in 8151 bytes as well. The differences
of time are within what this set-up resolves.
Maps of 1920x1080 pixels with 3 vector layers, each with another extent, requested at once: time until the
last one is answered.
| Requests at once | 1 process (default) | QFASTSERVER_WORKERS=4 |
|---|---|---|
| 1 | 60 ms | 64 ms |
| 4 | 209 ms | 79 ms |
| 8 | 435 ms | 154 ms |
| 16 | 792 ms | 281 ms |
With QFASTSERVER_GZIP=true the GetCapabilities document of the test project goes from 7681 to 1629 bytes
on the wire.
Machine: Intel Core Ultra 7 165H (16 cores, 22 threads), 31 GB of memory, Docker Engine 29.9 on WSL 2, laptop
on mains power. One run; the tables of the README were not measured again with this release.
Changes since the previous release
- Patch 0015: 32 bits PNG images which are empty or nearly empty are deflated with the run-length strategy
of zlib. In 3.44.15-qfs.1 an empty map took about 4.5 times the room it takes with the official image. QFASTSERVER_WORKERS: number of QGIS Server processes of the container. Default 1, as the official image.QFASTSERVER_GZIP: compression of the XML, JSON and text responses by the nginx of the image. Off by
default, as in the official image.- The command of the image is now
/usr/local/bin/qfastserver-start.sh. It applies the two options and
then runs/usr/local/bin/start-xvfb-nginx.sh, the start script of the official image, which is not
changed. A deployment which calls that script itself keeps working, without the two options.
See the README for the options and
UPSTREAM.md for the patches.
Known differences from upstream
- PNG files of vector maps are smaller; PNG files of pure imagery are about 20% larger (rows are written
without the filter search). - The image is built with the options of the upstream Debian packaging, without the desktop application and
the GRASS plugins, which the official server image does not contain either. - The command of the image is another script, which runs the one of the official image.
Not verified
- Python server plugins (for example
lizmap_server) with this image. - WFS transactions, GetPrint with an atlas, GeoPDF.
- The memory used by several processes with large projects, and
QFASTSERVER_WORKERSunder a load test.
QFastServer 3.44.15-qfs.1
First release of QFastServer for QGIS Server 3.44.15.
docker pull ghcr.io/envirosolutionspl/qfastserver:3.44.15-qfs.1Use the image exactly as qgis/qgis-server:3.44.15-noble: same entry point, ports, volumes and environment
variables.
Compatibility
| QGIS Server | 3.44.15 |
| Upstream tag and commit | final-3_44_15, 44b4ef051784192d6e4829dc008bca120e4d2388 |
| Patch series | 3.44, 14 patches, fork commit d3c624ee6ec6c706cf6b48564ce12a289494370d |
| Base image | qgis/qgis-server:3.44.15-noble |
| Qt | 5.15 |
Verification
- Responses compared with the official image
qgis/qgis-server:3.44.15-noble: 56 of 59 request types
identical. Not comparable, because they differ from run to run in the official image itself: 8 bits PNG,
a cascaded WMS layer as PNG, PDF. - Raster renderings compared with an unpatched build: 68 of 68 identical.
- Upstream tests (server, rasters, projects, vector layers, DXF): 139 of 139. They were run with the same
patches on the branchrelease-3_4428 commits after the tagfinal-3_44_15, not on the tag itself. - Upstream tests which need PostgreSQL: 36 of 39; the three failures do not depend on the patches (one is
on the upstream blocklist, one needs a PostgreSQL binary which the test container does not have, one
depends on the locale of the container).
Performance
The official image against this image, both started the same way (the nginx front of the image, 4 FastCGI
workers, parallel rendering off, 4 CPUs and 4 GB for the container, PostGIS 16 on the same host), k6, median
with one client and requests per second with 16 clients, no failed responses.
| Request | Median, 1 client | Requests/s, 16 clients |
|---|---|---|
| Tile 256x256, 1 vector layer, PNG | 10.9 → 4.5 ms | 304 → 459 |
| Tile 256x256, 3 vector layers, PNG | 11.0 → 7.9 ms | 222 → 271 |
| Map 1024x768, 1 vector layer, PNG | 30.3 → 18.0 ms | 70 → 95 |
| Map 1024x768, 3 vector layers, PNG | 44.6 → 27.5 ms | 42 → 57 |
| Map 1024x768, raster in the layer CRS, JPEG | 17.8 → 10.1 ms | 198 → 301 |
| Map 1024x768, raster reprojected, JPEG | 37.5 → 18.9 ms | 86 → 162 |
| GetFeatureInfo, JSON | 11.0 → 3.7 ms | 374 → 871 |
| GetLegendGraphic, PNG | 10.9 → 4.5 ms | 377 → 724 |
| WMTS GetTile | 10.9 → 5.3 ms | 261 → 356 |
| WFS GetFeature, GML | 44.6 → 37.0 ms | 57 → 74 |
Machine: Intel Core Ultra 7 165H (16 cores, 22 threads), 31 GB of memory, Docker Engine 29.9 on WSL 2.
The method and its limits are described in the README.
Changes
All 14 patches of the series; see UPSTREAM.md.
Known differences from upstream
- PNG files of vector maps are smaller; PNG files of pure imagery are about 20% larger (rows are written
without the filter search). - The image is built with the options of the upstream Debian packaging, without the desktop application and
the GRASS plugins, which the official server image does not contain either.
Not verified
- Python server plugins (for example
lizmap_server) with this image. - WFS transactions, GetPrint with an atlas, GeoPDF.