On Dec 2, 2019, at 3:02 AM, avinash.jha2493 <avinash.jha2...@gmail.com> wrote: > > The issue appears similar to what was mentioned here: > https://github.com/tensorflow/tensorboard/issues/1467
Yes, and OP posted a reply to his own issue a few days later giving the correct answer: TensorFlow needs to pick one of the two versions of SQLite and use it exclusively within that process. This is not a SQLite issue, it’s a SQLite *usage* issue. You can’t have two different versions of a library both using the same RAM data structures but with different assumptions about their binary layout. They’ve known about this problem for over a year, and they’re not moving on it. Apparently they want someone to contribute a fix for it. Maybe you? > Could someone guide on the next steps, have been playing with pragmas for a > while now, but didn't succeed. It’s not a configuration issue. They’re basically begging for data corruption with that setup. Mind, SQLite does allow multiple versions to access the same DB from *different* processes. The problem in TensorFlow is that they’ve got two different versions linked into the same binary, and both are trying to access the same DB. To avoid problems with this, you’d basically have to get rid of all globals in SQLite. If I’m interpreting the GCC map file correctly, there’s about 32 of them in recent versions of SQLite. For two versions of *any* library — not just SQLite — to be linked to the same program and work correctly together, they’d both have to have the same binary interpretation for each global variable, at the least. _______________________________________________ sqlite-users mailing list sqlite-users@mailinglists.sqlite.org http://mailinglists.sqlite.org/cgi-bin/mailman/listinfo/sqlite-users
