Huge reschedule-backlogs - after re-optimize, update

I recently updated to 24.06.02, re-optimized and rescheduled – as I’ve done several times over the past 8 months using FSRS – and I am surprised to see an unexpectedly huge backlog. I am unable to answer for myself the sorts of questions I frequently answer for others. :sweat_smile: Is this within the realm of normal? Am I doing something wrong?

TL;DR:

  • Why am I getting a huge reschedule-backlog? 400 cards after a normal re-optimize, and then 900 cards instead after updating versions (and a re-optimize) a week later?
  • Is FSRS in 24.06.02 so different from 23.12.1 that it accounts for this? What changed that would cause a 10% jump in my review history?
  • What does it mean that FSRS tells me it’s doing a great job fitting my review history – and has told me throughout the time I’ve been using it – if I’m getting a sudden backlog like this? Am I really so unpredictable?
  • Should I expect to keep getting a huge backlog every time I update?

My optimization “habit” –

Since I started using FSRS, every month or 2 – whenever I remember, or if I make a significant change (like changing what decks are in a preset) – I evaluate my current parameters, re-optimize, and re-evaluate. A little before-and-after check for how things are going.

Details

There has only been one time when my RMSE got worse after re-optimization. (I didn’t switch to the new parameters then, and I know that has been fixed to not be an issue again.) After getting new parameters, I usually reschedule, but I’m not a fanatic about it.

When I initially started using custom-code FSRS last year, I rescheduled-all using the Helper add-on (this was back before the recommendation against doing that!) – and I was rewarded with a backlog of about 400 cards. It was miserable for 2 months. Since then, my typical reschedule-backlog might be about 30 cards.

On a completely unrelated cycle, I periodically reschedule-all with the add-on to get load balancing. This is probably once a month or so, or after I have had some especially uneven days (incomplete Reviews, backlog catch-up, etc.). While I keep a careful record of my optimizations / parameters / evaluates (shocking, I’m sure) – I don’t really make a note of my reschedule-alls.


My last “typical” re-optimize:

4-15-24 using 23.12.1
evaluated 56K reviews at RMSE 1.72%
re-optimized, evaluated at RMSE 1.57%

Parameters

0.3050, 1.4696, 5.2671, 18.7406, 5.3726, 1.4943, 0.8948, 0.0241, 2.0988, 0.2850, 1.4503, 2.3259, 0.0354, 0.2415, 1.8678, 0.1807, 2.8789

I likely did a reschedule-all then; I can’t be sure. However, I know for sure that on 6-5-24, I did a reschedule-all to demonstrate Load Balance for someone, and there wasn’t much change. I can’t remember if I left that in place or undid it.


My next re-optimize:

6-24-24 using 23.12.1
evaluated 60K reviews at RMSE 1.49%
re-optimized, evaluated at RMSE 1.28%

Parameters

0.3212, 1.4901, 5.4759, 20.6647, 5.5531, 1.3864, 0.9140, 0.0200, 1.9740, 0.2504, 1.3648, 2.3066, 0.0528, 0.2716, 1.8885, 0.0842, 2.7771

Odd thing #1

After that, I ran a reschedule-all with the add-on, and it surprised me with a backlog of over 400 cards. That’s big for my collection (6% of my active cards), and more than a week’s workload. This was unexpected to me after using FSRS for 8+ months (the biggest reschedule-backlog since I first enabled FSRS), and with nothing significant changing in the past 2 months. I undid that reschedule and thought about what to do next.


The current situation:

A few days later on 6-27-24, I upgraded to 24.06.02. And then today, I took another run at re-optimizing.

7-1-24 using 24.06.02
evaluated 67K reviews at RMSE 2.28%
re-optimized, evaluated at 2.05%

Parameters

0.2712, 1.2870, 4.5993, 36.3857, 5.5666, 1.8591, 1.0206, 0.0141, 2.1699, 0.3620, 1.6549, 2.3402, 0.0826, 0.2489, 2.6217, 0.1198, 2.7451

Odd thing #2

Evaluated 67K reviews?! There’s no way I did ~7000 reviews in a week! I checked and I only studied 386 cards – FSRS Stats is counting 390 reps, main Stats is counting 600 reps. There were no major changes to my collection, the same 6 decks are in this preset, and only 3 of those have cards in them. I edited 100-ish notes, added a few others, and likely suspended / unsuspended some cards.

Odd thing #3

RMSE went the wrong way, to 2.28%. I guess a new version of FSRS can come with significant changes. But I optimized just a week ago, so that seems like a lot of change.

After the re-optimize, my new parameters shifted more-than-usual (by visual approximation) – not entirely expected with a different FSRS version, it seems like a lot after such a short time. The RMSE improved a little, but not as good as I expected from before updating.

Odd thing #4

Ran reschedule-all, and got a backlog of over 900 cards! What?! How is it possible that almost 14% of my active cards needed to be rescheduled?

Some notes on Stats

My desired retention has been 92% throughout – steady with what I was maintaining using SM-2. True retention has dropped a bit below that the past months or so, to ~91%.

Taking a look at my other Stats, with the current re-optimize, my average stability (for 95%) dropped from 1.49y to 1.28y. Then with the reschedule, my average interval (for 95%) had a commensurate drop from 1.26y to 1.03y.

1 Like

@Expertium Do you see what I mean about it being in the realm of your earlier issue, but not quite the same? I didn’t want to re-open an old bug if it wasn’t really related, but if you and @L.M.Sherlock would prefer, I can move the discussion over there. Thanks!

I fixed a dozen of bugs related with FSRS during 23.12.1-24.06.02. The version of FSRS-rs bumped from 0.0.0 to 0.6.4. So it’s really hard to list the difference between 24.06.02 and 23.12.1.

1 Like

The calculation method of RMSE is also changed. It’s common that the new RMSE is higher than the old RMSE.

1 Like

Could you share your collection? I need to analyze this problem deeply.

One thing I will mention is intervals of very high stability cards changing affects “average interval” a lot. (what is for 95 percent? add-on stats? for me, this changes nothing in the anki stats page)

Am I really so unpredictable?

A better place to look at is card info of the cards backlogged and how much change they had in case you’re only looking at the averages for this.

All of this might also just be a bug if there was an increase of 7K reviews “in a week” or “after an update”.

Happy to! I’ll message you.

I was referring to standard Stats – where the default is tracking the first 95% of the collection for long-tail stats like that. However it appears that the average is computed on the entire collection, regardless of %-age of the collection selected, so you can disregard that.

I’m always willing to look at Card Info, but it’s nearly 1000 cards. So Stats are going to have to offer a good enough insight until there’s a sign of what sorts of indicators I would be looking for on a single card.