Instant ALTER TABLE is not crash-safe

I'm using the official MariaDB mariadb:10.4.6-bionic image and I have a problem: my tables corrupt themselves. This had happened at least four times this month:

• on two different dev machines using docker-compose
• on the staging machine using docker swarm

Everytime the symptom is that a table exists in the information schema but it can't be read or altered at all. Any operation leads to a 'Table does not exists in engine' error.

The container logs contain this error at startup:

[ERROR] InnoDB: Table sfdb.specialist contains unrecognizable instant ALTER metadata

I've checked the server system variables with

SHOW VARIABLES;

And alter_algorithm has the value DEFAULT which means (if I understand the documentation correctly) that it should not use the INSTANT algorithm, only INPLACE and COPY.
https://mariadb.com/kb/en/library/server-system-variables/#alter_algorithm

Trying to dig into the MariaDB codebase led me to this check:
https://github.com/MariaDB/server/blob/cccfa9dcfe2c161779824c01a84edfa64fc4378a/storage/innobase/btr/btr0cur.cc#L465

But I do not have enough knowledge of the inner workings of MariaDB to understand what is going on.

What am I doing wrong? Should I set some flag or configuration to have a stable database in docker?

PS: if this JIRA is not the correct place to ask this can you point me to a more suitable place to post my question ?