This commit is contained in:
Daniel Pechersky 2025-09-17 14:51:58 +08:00 committed by GitHub
commit c086b06a67
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 281 additions and 114 deletions

View file

@ -243,6 +243,7 @@ Lee Doughty <32392044+leedoughty@users.noreply.github.com>
memchr <memchr@proton.me>
Max Romanowski <maxr777@proton.me>
Aldlss <ayaldlss@gmail.com>
Daniel Pechersky <danny.pechersky@gmail.com>
********************

12
Cargo.lock generated
View file

@ -124,6 +124,7 @@ dependencies = [
"once_cell",
"pbkdf2",
"percent-encoding-iri",
"permutation",
"phf 0.11.3",
"pin-project",
"prettyplease",
@ -2233,9 +2234,8 @@ dependencies = [
[[package]]
name = "fsrs"
version = "5.1.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "04954cc67c3c11ee342a2ee1f5222bf76d73f7772df08d37dc9a6cdd73c467eb"
version = "5.2.0"
source = "git+https://github.com/open-spaced-repetition/fsrs-rs.git#1e271981367454468391f1c686af03a0aa7aab3c"
dependencies = [
"burn",
"itertools 0.14.0",
@ -4561,6 +4561,12 @@ name = "percent-encoding-iri"
version = "2.2.0"
source = "git+https://github.com/ankitects/rust-url.git?rev=bb930b8d089f4d30d7d19c12e54e66191de47b88#bb930b8d089f4d30d7d19c12e54e66191de47b88"
[[package]]
name = "permutation"
version = "0.4.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "df202b0b0f5b8e389955afd5f27b007b00fb948162953f1db9c70d2c7e3157d7"
[[package]]
name = "pest"
version = "2.8.1"

View file

@ -33,8 +33,8 @@ git = "https://github.com/ankitects/linkcheck.git"
rev = "184b2ca50ed39ca43da13f0b830a463861adb9ca"
[workspace.dependencies.fsrs]
version = "5.1.0"
# git = "https://github.com/open-spaced-repetition/fsrs-rs.git"
# version = "5.1.0"
git = "https://github.com/open-spaced-repetition/fsrs-rs.git"
# path = "../open-spaced-repetition/fsrs-rs"
[workspace.dependencies]
@ -99,6 +99,7 @@ num_cpus = "1.17.0"
num_enum = "0.7.3"
once_cell = "1.21.3"
pbkdf2 = { version = "0.12", features = ["simple"] }
permutation = "0.4.1"
phf = { version = "0.11.3", features = ["macros"] }
pin-project = "1.1.10"
prettyplease = "0.2.34"

View file

@ -76,6 +76,7 @@ num_enum.workspace = true
once_cell.workspace = true
pbkdf2.workspace = true
percent-encoding-iri.workspace = true
permutation.workspace = true
phf.workspace = true
pin-project.workspace = true
prost.workspace = true

View file

@ -9,6 +9,7 @@ use fsrs::MemoryState;
use fsrs::FSRS;
use fsrs::FSRS5_DEFAULT_DECAY;
use fsrs::FSRS6_DEFAULT_DECAY;
use itertools::Either;
use itertools::Itertools;
use super::params::ignore_revlogs_before_ms_from_config;
@ -58,6 +59,18 @@ pub(crate) struct UpdateMemoryStateEntry {
pub ignore_before: TimestampMillis,
}
trait ChunkIntoVecs<T> {
fn chunk_into_vecs(&mut self, chunk_size: usize) -> impl Iterator<Item = Vec<T>>;
}
impl<T> ChunkIntoVecs<T> for Vec<T> {
fn chunk_into_vecs(&mut self, chunk_size: usize) -> impl Iterator<Item = Vec<T>> {
std::iter::from_fn(move || {
(!self.is_empty()).then(|| self.drain(..chunk_size.min(self.len())).collect())
})
}
}
impl Collection {
/// For each provided set of params, locate cards with the provided search,
/// and update their memory state.
@ -79,66 +92,114 @@ impl Collection {
let search =
SearchBuilder::all([search.into(), SearchNode::State(StateKind::New).negated()]);
let revlog = self.revlog_for_srs(search)?;
let reschedule = req.as_ref().map(|e| e.reschedule).unwrap_or_default();
let last_revlog_info = if reschedule {
Some(get_last_revlog_info(&revlog))
} else {
None
let Some(req) = &req else {
let items = fsrs_items_for_memory_states(
&FSRS::new(Some([].as_slice()))?,
revlog,
timing.next_day_at,
0.9,
ignore_before,
)?;
let on_updated_card = self.create_progress_closure(items.len())?;
// clear FSRS data if FSRS is disabled
self.clear_fsrs_data_for_cards(
items.into_iter().map(|(card_id, _)| card_id),
usn,
on_updated_card,
)?;
continue;
};
let mut rescheduler = self
.get_config_bool(BoolKey::LoadBalancerEnabled)
.then(|| Rescheduler::new(self))
.transpose()?;
let fsrs = FSRS::new(req.as_ref().map(|w| &w.params[..]).or(Some([].as_slice())))?;
let decay = req.as_ref().map(|w| get_decay_from_params(&w.params));
let historical_retention = req.as_ref().map(|w| w.historical_retention);
let fsrs = FSRS::new(Some(&req.params[..]))?;
let last_revlog_info = req.reschedule.then(|| get_last_revlog_info(&revlog));
let items = fsrs_items_for_memory_states(
&fsrs,
revlog,
timing.next_day_at,
historical_retention.unwrap_or(0.9),
req.historical_retention,
ignore_before,
)?;
let mut progress = self.new_progress_handler::<ComputeMemoryProgress>();
progress.update(false, |s| s.total_cards = items.len() as u32)?;
for (idx, (card_id, item)) in items.into_iter().enumerate() {
progress.update(true, |state| state.current_cards = idx as u32 + 1)?;
let mut card = self.storage.get_card(card_id)?.or_not_found(card_id)?;
let original = card.clone();
if let Some(req) = &req {
let preset_desired_retention = req.preset_desired_retention;
let mut on_updated_card = self.create_progress_closure(items.len())?;
let (items, cards_without_items): (Vec<(CardId, FsrsItemForMemoryState)>, Vec<CardId>) =
items.into_iter().partition_map(|(card_id, item)| {
if let Some(item) = item {
Either::Left((card_id, item))
} else {
Either::Right(card_id)
}
});
let decay = get_decay_from_params(&req.params);
// Store decay and desired retention in the card so that add-ons, card info,
// stats and browser search/sorts don't need to access the deck config.
// Unlike memory states, scheduler doesn't use decay and dr stored in the card.
let set_decay_and_desired_retention = move |card: &mut Card| {
let deck_id = card.original_or_current_deck_id();
let desired_retention = *req
.deck_desired_retention
.get(&deck_id)
.unwrap_or(&preset_desired_retention);
.unwrap_or(&req.preset_desired_retention);
card.desired_retention = Some(desired_retention);
card.decay = decay;
if let Some(item) = item {
card.set_memory_state(&fsrs, Some(item), historical_retention.unwrap())?;
// if rescheduling
if let Some(reviews) = &last_revlog_info {
// and we have a last review time for the card
if let Some(last_info) = reviews.get(&card.id) {
if let Some(last_review) = &last_info.last_reviewed_at {
let days_elapsed =
timing.next_day_at.elapsed_days_since(*last_review) as i32;
// and the card's not new
if let Some(state) = &card.memory_state {
// or in (re)learning
if card.ctype == CardType::Review {
let deck = self
card.decay = Some(decay);
};
self.update_memory_state_for_itemless_cards(
cards_without_items,
set_decay_and_desired_retention,
usn,
&mut on_updated_card,
)?;
let mut rescheduler = self
.get_config_bool(BoolKey::LoadBalancerEnabled)
.then(|| Rescheduler::new(self))
.transpose()?;
let reschedule = move |card: &mut Card,
collection: &mut Self,
fsrs: &FSRS|
-> Result<()> {
// we are rescheduling
let Some(last_revlog_info) = &last_revlog_info else {
return Ok(());
};
// we have a last review time for the card
let Some(last_info) = last_revlog_info.get(&card.id) else {
return Ok(());
};
let Some(last_review) = &last_info.last_reviewed_at else {
return Ok(());
};
// the card isn't in (re)learning
if card.ctype != CardType::Review {
return Ok(());
};
let deck = collection
.get_deck(card.original_or_current_deck_id())?
.or_not_found(card.original_or_current_deck_id())?;
let deckconfig_id = deck.config_id().unwrap();
// reschedule it
let days_elapsed = timing.next_day_at.elapsed_days_since(*last_review) as i32;
let original_interval = card.interval;
let interval = fsrs.next_interval(
Some(state.stability),
desired_retention,
Some(
card.memory_state
.expect("We set it before this function is called")
.stability,
),
card.desired_retention
.expect("We set it before this function is called"),
0,
);
card.interval = rescheduler
@ -150,53 +211,128 @@ impl Collection {
req.max_interval,
days_elapsed as u32,
deckconfig_id,
get_fuzz_seed(&card, true),
get_fuzz_seed(card, true),
)
})
.unwrap_or_else(|| {
with_review_fuzz(
card.get_fuzz_factor(true),
interval,
1,
req.max_interval,
)
with_review_fuzz(card.get_fuzz_factor(true), interval, 1, req.max_interval)
});
let due = if card.original_due != 0 {
&mut card.original_due
} else {
&mut card.due
};
let new_due = (timing.days_elapsed as i32)
- days_elapsed
+ card.interval as i32;
let new_due = (timing.days_elapsed as i32) - days_elapsed + card.interval as i32;
if let Some(rescheduler) = &mut rescheduler {
rescheduler.update_due_cnt_per_day(
*due,
new_due,
deckconfig_id,
);
rescheduler.update_due_cnt_per_day(*due, new_due, deckconfig_id);
}
*due = new_due;
// Add a rescheduled revlog entry
self.log_rescheduled_review(
&card,
original_interval,
collection.log_rescheduled_review(card, original_interval, usn)?;
Ok(())
};
self.update_memory_state_for_cards_with_items(
items,
&fsrs,
set_decay_and_desired_retention,
reschedule,
usn,
on_updated_card,
)?;
}
Ok(())
}
fn create_progress_closure(&self, item_count: usize) -> Result<impl FnMut() -> Result<()>> {
let mut progress = self.new_progress_handler::<ComputeMemoryProgress>();
progress.update(false, |s| {
s.total_cards = item_count as u32;
s.current_cards = 1;
})?;
let on_updated_card = move || progress.update(true, |p| p.current_cards += 1);
Ok(on_updated_card)
}
}
}
} else {
// clear memory states if item is None
card.memory_state = None;
}
} else {
// clear FSRS data if FSRS is disabled
fn clear_fsrs_data_for_cards(
&mut self,
cards: impl Iterator<Item = CardId>,
usn: Usn,
mut on_updated_card: impl FnMut() -> Result<()>,
) -> Result<()> {
for card_id in cards {
let mut card = self.storage.get_card(card_id)?.or_not_found(card_id)?;
let original = card.clone();
card.clear_fsrs_data();
}
self.update_card_inner(&mut card, original, usn)?;
on_updated_card()?
}
Ok(())
}
fn update_memory_state_for_itemless_cards(
&mut self,
cards: Vec<CardId>,
mut set_decay_and_desired_retention: impl FnMut(&mut Card),
usn: Usn,
mut on_updated_card: impl FnMut() -> Result<()>,
) -> Result<()> {
for card_id in cards {
let mut card = self.storage.get_card(card_id)?.or_not_found(card_id)?;
let original = card.clone();
set_decay_and_desired_retention(&mut card);
card.memory_state = None;
self.update_card_inner(&mut card, original, usn)?;
on_updated_card()?;
}
Ok(())
}
fn update_memory_state_for_cards_with_items(
&mut self,
items: Vec<(CardId, FsrsItemForMemoryState)>,
fsrs: &FSRS,
mut set_decay_and_desired_retention: impl FnMut(&mut Card),
mut maybe_reschedule_card: impl FnMut(&mut Card, &mut Self, &FSRS) -> Result<()>,
usn: Usn,
mut on_updated_card: impl FnMut() -> Result<()>,
) -> Result<()> {
const FSRS_BATCH_SIZE: usize = 1000;
let mut to_update = Vec::new();
let mut fsrs_items = Vec::new();
let mut starting_states = Vec::new();
for (card_id, item) in items.into_iter() {
to_update.push(card_id);
fsrs_items.push(item.item);
starting_states.push(item.starting_state);
}
// fsrs.memory_state_batch is O(nm) where n is the number of cards and m is the
// max review count between all items. Therefore we want to pass batches
// to fsrs.memory_state_batch where the review count is relatively even.
let mut p = permutation::sort_unstable_by_key(&fsrs_items, |item| item.reviews.len());
p.apply_slice_in_place(&mut to_update);
p.apply_slice_in_place(&mut fsrs_items);
p.apply_slice_in_place(&mut starting_states);
for ((to_update, fsrs_items), starting_states) in to_update
.chunk_into_vecs(FSRS_BATCH_SIZE)
.zip_eq(fsrs_items.chunk_into_vecs(FSRS_BATCH_SIZE))
.zip_eq(starting_states.chunk_into_vecs(FSRS_BATCH_SIZE))
{
let memory_states = fsrs.memory_state_batch(fsrs_items, starting_states)?;
for (card_id, memory_state) in to_update.into_iter().zip_eq(memory_states) {
let mut card = self.storage.get_card(card_id)?.or_not_found(card_id)?;
let original = card.clone();
set_decay_and_desired_retention(&mut card);
card.memory_state = Some(memory_state.into());
maybe_reschedule_card(&mut card, self, fsrs)?;
self.update_card_inner(&mut card, original, usn)?;
on_updated_card()?;
}
}
Ok(())
@ -269,7 +405,7 @@ impl Card {
}
}
#[derive(Debug)]
#[derive(Debug, Clone)]
pub(crate) struct FsrsItemForMemoryState {
pub item: FSRSItem,
/// When revlogs have been truncated, this stores the initial state at first
@ -504,4 +640,26 @@ mod tests {
);
Ok(())
}
mod update_memory_state {
use super::*;
use crate::collection::CollectionBuilder;
#[test]
fn smoke() {
let mut collection = CollectionBuilder::default().build().unwrap();
let entry = UpdateMemoryStateEntry {
req: None,
search: SearchNode::WholeCollection,
ignore_before: TimestampMillis(0),
};
collection
.transact(Op::UpdateDeckConfig, |collection| {
collection.update_memory_state(vec![entry]).unwrap();
Ok(())
})
.unwrap();
}
}
}