CCC's rcwatcher task has been tagging fixed images as corrupt (though not updating the database entries). As a result, I have shutdown the bot entirely pending investigation.
Some investigation revealed that the redis queue is currently 490,252 entries long and currently still processing uploads from Feb 12.
The crashes really hurt the bot's effectiveness. A new check needs to be added for cases like this and -- ideally -- an automated process developed to spin up (and down) workers dynamically.