[gdal-dev] [EXTERNAL] Re: gdal.Rasterize with same OGR dataset from two python threads

[Meyer, Jesse R. (GSFC-618.0)[SCIENCE SYSTEMS AND APPLICATIONS INC]]

How can I “open” a handle to a pre-existing memory dataset?  That sounds like it may work for me.

As a matter of semantics, my sense of what GetNextFeature() would return would be a local view of the database on a per thread basis.  Each thread would have its own cursor into the database, said another way.

Best,
Jesse

Lead Computer Scientist
Science Systems and Applications, Inc.
Dr Compton Tucker Team
NASA Goddard Space Flight Center

From: Even Rouault <even.rouault at spatialys.com>
Date: Monday, October 28, 2024 at 12:08 PM
To: Meyer, Jesse R. (GSFC-618.0)[SCIENCE SYSTEMS AND APPLICATIONS INC] <jesse.r.meyer at nasa.gov>, Meyer, Jesse R. (GSFC-618.0)[SCIENCE SYSTEMS AND APPLICATIONS INC] via gdal-dev <gdal-dev at lists.osgeo.org>
Subject: [EXTERNAL] Re: [gdal-dev] gdal.Rasterize with same OGR dataset from two python threads
CAUTION: This email originated from outside of NASA.  Please take care when clicking links or opening attachments.  Use the "Report Message" button to report suspicious messages to the NASA SOC.




Le 28/10/2024 à 17:01, Meyer, Jesse R. (GSFC-618.0)[SCIENCE SYSTEMS AND APPLICATIONS INC] via gdal-dev a écrit :
I have two calls to gdal.Rasterize, each of which target a separate GDAL memory dataset but source the same OGR memory dataset, that I hoped could be ran in parallel using Python’s concurrent futures.  The idea being that each GDAL call unlocks the Python GIL, and performing read only operations on the vector database (except for storing memory for the results) could in principle be a safe and effective optimization, as the feature layers themselves are not mutated.  The SQL dialect is SQLite, so presumably the OGR dataset has to be converted to a SQLite (memory) database.  Technically SQLite supports multiple readers just fine, but this doesn’t mean GDAL/OGR does.  The multithreading documentation page doesn’t explicitly mention OGR / vector datasets but I presume they inherit similar stateful restrictions (Yes RFC 101 is coming).  However, running these SQL queries at the same times causes OGR to trip over itself (I presume OGR assumes only one query statement is being evaluated at the same time).

So I think the intended work around is either: accept this is as a serially dependent task, or copy the dataset and have each Rasterize() work on a copy, yes?
I'm not clear if you use the same Python source vector dataset, or if you open your source dataset once for each thread ?  The first case is a big no no: anything could happen, including wrong results and crashes. One object per thread is the way to go. If the processing is very intensive on acquiring source features, you may hit a global lock at the SQLite level, but there isn't much we can do about that. Or you need to use multi-processing parallelization instead of multi-threading. But you certainly don't need to copy your source dataset.


In the same spirit as RFC 101, which gives some thread safety to raster read-only workloads, is there interest in expanding this to vector datasets?

That would be tricky. What would be the expect result if a user would use GetNextFeature() on a thread-safe OGRLayer...: would users expect each thread to see all features or features would be distributed among calling threads ?
Even

--

http://www.spatialys.com<http://www.spatialys.com/>

My software is free, but my time generally not.

Butcher of all kinds of standards, open or closed formats. At the end, this is just about bytes.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/gdal-dev/attachments/20241028/7bc5fcd7/attachment.htm>