Pilar

Keep a file with each error message and a one line note about what caused it. After thirty entries you have your own index and you stop searching for them.

Yes. It is a measurement, not a punishment.

Completely expected, and the software is doing exactly what it should.

The scheduler aims for a target retention, often around 90 percent. That means one card in ten is supposed to fail at the moment it comes due. If nothing ever failed, your intervals would be too short and you would be spending your life reviewing things you already know. A card that survived to a nine month interval and then died has bought you eight months of not thinking about it.

What would worry me is a pattern: same card, dies at every long interval, twelve times running. That is a leech and the card is badly written. One-off blanks are the price of long intervals.

Ask them what decision they make with it. If nobody can name one, it can run daily.

You have built recognition cards, not recall cards. The giveaway is that the surrounding sentence uniquely identifies the answer before you have thought about anything.

Two cheap fixes that do not need a full rewrite:

• Cut the tail of the sentence that gives the game away. If the card reads The [...] nerve controls the muscles of facial expression, the second half is the whole answer.
• One card per fact, phrased as the question you would actually be asked.

I would rewrite the 60 cards you care about most and let the rest die. Sunk cost on flashcards is real and expensive.

The money is already spent. The only question is whether you keep paying rent on them every morning.

Two million rows a night is small enough that full partition rebuilds are affordable. Take the simple correctness win while you can still afford it.

Turn the daily limit off for a moment and look at what the 912 actually are. If 600 of them are cards you have seen twenty times and always pass, you can honestly push them out to a longer interval. If 600 of them are cards you have never once got right, that is a card design problem and no scheduling trick fixes it.

And that thin tail is very often a single mobile app version with a bad offline queue.

Do you fail the run on a schema change, or warn and carry on with the columns you know?

We had no alert at all, which is how it became an incident rather than a chore.

Two out of three. We found out from a person the first time and that was easily the worst one.

Whichever you choose, snapshot the dimension daily if history matters at all. Rebuilds quietly destroy your ability to answer what something looked like in March.