In the test case above, upon restart InnoDB writes "failed to read" error messages into the log:
However, the table is readable.
The effect is not necessarily limited to MTR. These are the elements of the use case which I couldn't get rid of while reproducing outside MTR:
- server had to be bootstrapped without InnoDB;
- server had to be run with performance schema;
- provider had to be installed into mysql.plugin rather than loaded via a startup option.
I'm not sure how important each factor is, possibly they just cause a "lucky" timing or sequence of events upon startup. But this way it happens to me nearly deterministically. In rare cases only the first error message is written.
The specific provider is not important.