wiki:rfc70_output_format_guess

RFC 70: Guessing output format from output file name extension for utilities

Author: Even Rouault
Contact: even.rouault@…
Started: Aug 2017
Status: Adopted, implemented
Implementation version: 2.3.0

Summary

This proposal is to add syntaxic sugar to make GDAL and OGR command line utilities, so they take into account the extension of the output filename to guess which output driver to use, when it is not explicitly specified with -f / -of switch.

Motivation

Currently command line utilities require to explicitly specify the output format when not wishing to use the default format (generally GeoTIFF for raster, and Shapefile for vector). But this is rather counter-intuitive. For example "gdal_translate in.tif out.png" will generate a GeoTIFF, and "ogr2ogr out.gpkg in.shp" a shapefile. So you have to specify respectively -of PNG and -f GPKG to get the expected result.

Guessing the output format from the extension of the output filename is for example a behaviour found in ImageMagick? convert utility, or in OpenJPEG opj_compress/opj_decompress utilities.

Changes in C/C++ and Python utilities

Command line utilities, when neither -f nor -of are specified (note: since r39878 both switches can be indifferently used), will loop through the registered drivers and check if one or several drivers, with output capabilities, declare to recognize the extension of the output filename.

  • When one and only one driver declares this extension (.tif, .png, .jpg ect), it will be used automatically
  • When several drivers declare this extension (for example KML and LIBKML for .kml), the utility will select the first registered driver (except netCDF instead of GMT for .nc files), and a warning is emitted specifying which driver is used
  • When no driver declares this extension, and the extension is not empty (e.g a .mpg filename), the utility will error out

For completness:

  • When there's no extension, and no prefix is recognized (see below), the default output driver will be silently used, as currently

Since at least GDAL 1.10, the base of this logic already exists since a warning is emitted for C/C++ utilities, when the extension of the output format is known to be recognized by another driver than the default output driver.

Similarly, for vector output, if doing something like "ogr2ogr PG:dbname=mydb out.shp", a PG:dbname=mydb directory is created with shapefiles, instead of ingesting the shapefile into PostgreSQL. A warning is emitted in that case since the PG driver declares the PG: prefix in its metadata. The new behaviour will be to imply the -update switch in such situation.

When the utilities are available as library functions (GDALTranslate(), etc.), output format guessing will also be applied if the -f/-of switch is not specified

Changes in SWIG bindings

For librarified utilities (gdal.Translate, etc.), the format argument now defaults to None.

Potential issues

There might be some fragility with the new logic in the situation where a GDAL version has only one driver that supports extension xxx, but a later version adds another driver that also supports extension xxx (or another distribution of the same version has a plugin that handles xxx). So scripts that did "gdal_translate in out.xxx" would now error out in the next version since several drivers are available.

Bottom line: always specify the output driver when reliability/reproducibility is desired.

This RFC mostly helps for interactive conversions where the less you type the better.

Backward compatibility

This will break scripts that use an output filename whose extension is matched by a driver which is not the default one. This incompatibility is rather unlikely since previous GDAL versions already emit a warning in this situation (for C/C++ utilities only. for Python utilities default driver is silently used), so people have likely specified the output driver if they really want to do "gdal_translate in.tif out.png -of GTiff".

MIGRATION_GUIDE.TXT will mention those potential caveats.

Testing

The existing autotest suite should continue to pass (with a few changes related to tests for the current behaviour)

Implementation

Implementation will be done by Even Rouault

Proposed implementation is in https://github.com/rouault/gdal2/tree/rfc70

Diff: https://github.com/OSGeo/gdal/compare/trunk...rouault:rfc70

Voting history

+1 from JukkaR, TamasS, DanielM and EvenR

Last modified 10 months ago Last modified on Nov 20, 2017 6:24:38 AM