federation: parallel sending

federation-send-parallel
phiresky 2024-04-13 23:18:28 +02:00
parent 3a0c1dca90
commit 539f06af97
4 changed files with 303 additions and 134 deletions

4
Cargo.lock generated
View File

@ -5439,9 +5439,9 @@ checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"
[[package]]
name = "tokio"
version = "1.36.0"
version = "1.37.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "61285f6515fa018fb2d1e46eb21223fff441ee8db5d0f1435e8ab4f5cdb80931"
checksum = "1adbebffeca75fcfd058afa480fb6c0b81e165a0323f9c9d39c9697e37c46787"
dependencies = [
"backtrace",
"bytes",

View File

@ -139,7 +139,7 @@ anyhow = { version = "1.0.81", features = [
diesel_ltree = "0.3.1"
typed-builder = "0.18.1"
serial_test = "2.0.0"
tokio = { version = "1.36.0", features = ["full"] }
tokio = { version = "1.37.0", features = ["full"] }
regex = "1.10.3"
once_cell = "1.19.0"
diesel-derive-newtype = "2.1.0"

View File

@ -73,19 +73,16 @@ async fn start_stop_federation_workers(
// create new worker
let config = federation_config.clone();
let stats_sender = stats_sender.clone();
let pool = pool.clone();
workers.insert(
instance.id,
CancellableTask::spawn(WORKER_EXIT_TIMEOUT, move |stop| {
let instance = instance.clone();
let req_data = config.clone().to_request_data();
let config = config.clone();
let stats_sender = stats_sender.clone();
let pool = pool.clone();
async move {
InstanceWorker::init_and_loop(
instance,
req_data,
&mut DbPool::Pool(&pool),
config,
stop,
stats_sender,
)

View File

@ -7,7 +7,7 @@ use crate::util::{
};
use activitypub_federation::{
activity_sending::SendActivityTask,
config::Data,
config::{Data, FederationConfig},
protocol::context::WithContext,
};
use anyhow::{Context, Result};
@ -22,22 +22,22 @@ use lemmy_db_schema::{
instance::{Instance, InstanceForm},
site::Site,
},
utils::{naive_now, DbPool},
utils::{naive_now, ActualDbPool, DbPool},
};
use lemmy_db_views_actor::structs::CommunityFollowerView;
use once_cell::sync::Lazy;
use reqwest::Url;
use std::{
collections::{HashMap, HashSet},
collections::{BinaryHeap, HashMap, HashSet},
ops::{Add, Deref},
time::Duration,
};
use tokio::{sync::mpsc::UnboundedSender, time::sleep};
use tokio::{
sync::mpsc::{self, UnboundedSender},
time::sleep,
};
use tokio_util::sync::CancellationToken;
/// Check whether to save state to db every n sends if there's no failures (during failures state is saved after every attempt)
/// This determines the batch size for loop_batch. After a batch ends and SAVE_STATE_EVERY_TIME has passed, the federation_queue_state is updated in the DB.
static CHECK_SAVE_STATE_EVERY_IT: i64 = 100;
/// Save state to db after this time has passed since the last state (so if the server crashes or is SIGKILLed, less than X seconds of activities are resent)
static SAVE_STATE_EVERY_TIME: Duration = Duration::from_secs(60);
/// interval with which new additions to community_followers are queried.
@ -57,6 +57,16 @@ static FOLLOW_ADDITIONS_RECHECK_DELAY: Lazy<chrono::TimeDelta> = Lazy::new(|| {
/// This is expected to happen pretty rarely and updating it in a timely manner is not too important.
static FOLLOW_REMOVALS_RECHECK_DELAY: Lazy<chrono::TimeDelta> =
Lazy::new(|| chrono::TimeDelta::try_hours(1).expect("TimeDelta out of bounds"));
static CONCURRENT_SENDS: Lazy<i64> = Lazy::new(|| {
std::env::var("LEMMY_FEDERATION_CONCURRENT_SENDS_PER_INSTANCE")
.ok()
.and_then(|s| s.parse().ok())
.unwrap_or(8)
});
/// Maximum number of successful sends to allow out of order
const MAX_SUCCESSFULS: usize = 1000;
pub(crate) struct InstanceWorker {
instance: Instance,
// load site lazily because if an instance is first seen due to being on allowlist,
@ -67,60 +77,116 @@ pub(crate) struct InstanceWorker {
site: Option<Site>,
followed_communities: HashMap<CommunityId, HashSet<Url>>,
stop: CancellationToken,
context: Data<LemmyContext>,
config: FederationConfig<LemmyContext>,
stats_sender: UnboundedSender<(String, FederationQueueState)>,
last_full_communities_fetch: DateTime<Utc>,
last_incremental_communities_fetch: DateTime<Utc>,
state: FederationQueueState,
last_state_insert: DateTime<Utc>,
pool: ActualDbPool,
}
#[derive(Debug, PartialEq, Eq)]
struct SendSuccessInfo {
activity_id: ActivityId,
published: Option<DateTime<Utc>>,
was_skipped: bool,
}
impl PartialOrd for SendSuccessInfo {
fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
other.activity_id.partial_cmp(&self.activity_id)
}
}
impl Ord for SendSuccessInfo {
fn cmp(&self, other: &Self) -> std::cmp::Ordering {
other.activity_id.cmp(&self.activity_id)
}
}
enum SendActivityResult {
Success(SendSuccessInfo),
Failure {
fail_count: i32,
// activity_id: ActivityId,
},
}
impl InstanceWorker {
pub(crate) async fn init_and_loop(
instance: Instance,
context: Data<LemmyContext>,
pool: &mut DbPool<'_>, // in theory there's a ref to the pool in context, but i couldn't get that to work wrt lifetimes
config: FederationConfig<LemmyContext>,
// pool: ActualDbPool, // in theory there's a ref to the pool in context, but i couldn't get that to work wrt lifetimes
stop: CancellationToken,
stats_sender: UnboundedSender<(String, FederationQueueState)>,
) -> Result<(), anyhow::Error> {
let state = FederationQueueState::load(pool, instance.id).await?;
let state =
FederationQueueState::load(&mut config.to_request_data().pool(), instance.id).await?;
let pool = config.to_request_data().inner_pool().clone();
let mut worker = InstanceWorker {
instance,
site_loaded: false,
site: None,
followed_communities: HashMap::new(),
stop,
context,
config,
stats_sender,
last_full_communities_fetch: Utc.timestamp_nanos(0),
last_incremental_communities_fetch: Utc.timestamp_nanos(0),
state,
last_state_insert: Utc.timestamp_nanos(0),
pool,
};
worker.loop_until_stopped(pool).await
worker.loop_until_stopped().await
}
/// loop fetch new activities from db and send them to the inboxes of the given instances
/// this worker only returns if (a) there is an internal error or (b) the cancellation token is cancelled (graceful exit)
pub(crate) async fn loop_until_stopped(
&mut self,
pool: &mut DbPool<'_>,
) -> Result<(), anyhow::Error> {
let save_state_every = chrono::Duration::from_std(SAVE_STATE_EVERY_TIME).expect("not negative");
self.update_communities(pool).await?;
async fn loop_until_stopped(&mut self) -> Result<()> {
self.initial_fail_sleep().await?;
let mut latest_id = self.get_latest_id().await?;
// activities that have been successfully sent but
// that are not the lowest number and thus can't be written to the database yet
let mut successfuls = BinaryHeap::<SendSuccessInfo>::new();
let mut in_flight: i64 = 0;
let (report_inbox_result, mut receive_inbox_result) =
tokio::sync::mpsc::unbounded_channel::<SendActivityResult>();
while !self.stop.is_cancelled() {
self.loop_batch(pool).await?;
if self.stop.is_cancelled() {
break;
// check if we need to wait for a send to finish before sending the next one
let need_wait_for_event = (in_flight != 0 && self.state.fail_count > 0)
|| successfuls.len() > MAX_SUCCESSFULS
|| in_flight >= *CONCURRENT_SENDS;
if need_wait_for_event || receive_inbox_result.len() > 4 {
self
.handle_send_results(&mut receive_inbox_result, &mut successfuls, &mut in_flight)
.await?;
} else {
self.update_communities().await?;
let last_successful_id = self
.state
.last_successful_id
.map(|e| e.0)
.expect("set above");
let next_id = ActivityId(last_successful_id + (successfuls.len() as i64) + in_flight + 1);
if next_id > latest_id {
latest_id = self.get_latest_id().await?;
if next_id > latest_id {
// no more work to be done, wait before rechecking
tokio::select! {
() = sleep(*WORK_FINISHED_RECHECK_DELAY) => {},
() = self.stop.cancelled() => {}
}
continue;
}
}
in_flight += 1;
self
.spawn_send_if_needed(next_id, report_inbox_result.clone())
.await?;
}
if (Utc::now() - self.last_state_insert) > save_state_every {
self.save_and_send_state(pool).await?;
}
self.update_communities(pool).await?;
}
// final update of state in db
self.save_and_send_state(pool).await?;
// final update of state in db on shutdown
self.save_and_send_state().await?;
Ok(())
}
@ -137,6 +203,11 @@ impl InstanceWorker {
return Ok(());
}
let remaining = required - elapsed;
tracing::debug!(
"{}: fail-sleeping for {:?} before starting queue",
self.instance.domain,
remaining
);
tokio::select! {
() = sleep(remaining) => {},
() = self.stop.cancelled() => {}
@ -144,78 +215,174 @@ impl InstanceWorker {
}
Ok(())
}
/// send out a batch of CHECK_SAVE_STATE_EVERY_IT activities
async fn loop_batch(&mut self, pool: &mut DbPool<'_>) -> Result<()> {
let latest_id = get_latest_activity_id(pool).await?;
let mut id = if let Some(id) = self.state.last_successful_id {
id
} else {
/// get newest activity id and set it as last_successful_id if it's the first time this instance is seen
async fn get_latest_id(&mut self) -> Result<ActivityId> {
let latest_id = get_latest_activity_id(&mut self.pool()).await?;
if let None = self.state.last_successful_id {
// this is the initial creation (instance first seen) of the federation queue for this instance
// skip all past activities:
self.state.last_successful_id = Some(latest_id);
// save here to ensure it's not read as 0 again later if no activities have happened
self.save_and_send_state(pool).await?;
latest_id
};
if id >= latest_id {
// no more work to be done, wait before rechecking
tokio::select! {
() = sleep(*WORK_FINISHED_RECHECK_DELAY) => {},
() = self.stop.cancelled() => {}
self.save_and_send_state().await?;
}
Ok(latest_id)
}
async fn handle_send_results(
&mut self,
receive_inbox_result: &mut mpsc::UnboundedReceiver<SendActivityResult>,
successfuls: &mut BinaryHeap<SendSuccessInfo>,
in_flight: &mut i64,
) -> Result<(), anyhow::Error> {
let force_write = false;
let mut events = Vec::new();
// wait for at least one event but if there's multiple handle them all
receive_inbox_result.recv_many(&mut events, 1000).await;
for event in events {
match event {
SendActivityResult::Success(s) => {
self.state.fail_count = 0;
*in_flight -= 1;
if !s.was_skipped {
self.mark_instance_alive().await?;
}
successfuls.push(s);
}
SendActivityResult::Failure { fail_count, .. } => {
if fail_count > self.state.fail_count {
// override fail count - if multiple activities are currently sending this value may get conflicting info but that's fine
self.state.fail_count = fail_count;
self.state.last_retry = Some(Utc::now());
}
}
}
}
self
.pop_successfuls_and_write(successfuls, force_write)
.await?;
Ok(())
}
async fn mark_instance_alive(&mut self) -> Result<()> {
// Activity send successful, mark instance as alive if it hasn't been updated in a while.
let updated = self.instance.updated.unwrap_or(self.instance.published);
if updated.add(Days::new(1)) < Utc::now() {
self.instance.updated = Some(Utc::now());
let form = InstanceForm::builder()
.domain(self.instance.domain.clone())
.updated(Some(naive_now()))
.build();
Instance::update(&mut self.pool(), self.instance.id, form).await?;
}
Ok(())
}
/// checks whether the highest successful id can be updated and writes to db if so
async fn pop_successfuls_and_write(
&mut self,
successfuls: &mut BinaryHeap<SendSuccessInfo>,
force_write: bool,
) -> Result<()> {
let Some(mut last_id) = self.state.last_successful_id else {
tracing::warn!("should be impossible: last successful id is None");
return Ok(());
};
tracing::debug!(
"last: {:?}, next: {:?}, currently in successfuls: {:?}",
last_id,
successfuls.peek(),
successfuls.iter()
);
while successfuls
.peek()
.map(|a| &a.activity_id == &ActivityId(last_id.0 + 1))
.unwrap_or(false)
{
let next = successfuls.pop().unwrap();
last_id = next.activity_id;
self.state.last_successful_id = Some(next.activity_id);
self.state.last_successful_published_time = next.published;
}
let save_state_every = chrono::Duration::from_std(SAVE_STATE_EVERY_TIME).expect("not negative");
if force_write || (Utc::now() - self.last_state_insert) > save_state_every {
self.save_and_send_state().await?;
}
Ok(())
}
async fn spawn_send_if_needed(
&mut self,
activity_id: ActivityId,
report: UnboundedSender<SendActivityResult>,
) -> Result<()> {
let Some(ele) = get_activity_cached(&mut self.pool(), activity_id)
.await
.context("failed reading activity from db")?
else {
tracing::debug!("{}: {:?} does not exist", self.instance.domain, activity_id);
report.send(SendActivityResult::Success(SendSuccessInfo {
activity_id,
published: None,
was_skipped: true,
}))?;
return Ok(());
};
let activity = &ele.0;
let inbox_urls = self
.get_inbox_urls(activity)
.await
.context("failed figuring out inbox urls")?;
if inbox_urls.is_empty() {
tracing::debug!("{}: {:?} no inboxes", self.instance.domain, activity.id);
report.send(SendActivityResult::Success(SendSuccessInfo {
activity_id,
published: Some(activity.published),
was_skipped: true,
}))?;
return Ok(());
}
let mut processed_activities = 0;
while id < latest_id
&& processed_activities < CHECK_SAVE_STATE_EVERY_IT
&& !self.stop.is_cancelled()
{
id = ActivityId(id.0 + 1);
processed_activities += 1;
let Some(ele) = get_activity_cached(pool, id)
.await
.context("failed reading activity from db")?
else {
tracing::debug!("{}: {:?} does not exist", self.instance.domain, id);
self.state.last_successful_id = Some(id);
continue;
};
if let Err(e) = self.send_retry_loop(pool, &ele.0, &ele.1).await {
let inbox_urls = inbox_urls.into_iter().collect();
let initial_fail_count = self.state.fail_count;
let data = self.config.to_request_data();
let stop = self.stop.clone();
let domain = self.instance.domain.clone();
tokio::spawn(async move {
if let Err(e) = InstanceWorker::send_retry_loop(
&ele.0,
&ele.1,
inbox_urls,
report,
initial_fail_count,
domain,
data,
stop,
)
.await
{
tracing::warn!(
"sending {} errored internally, skipping activity: {:?}",
ele.0.ap_id,
e
);
}
if self.stop.is_cancelled() {
return Ok(());
}
// send success!
self.state.last_successful_id = Some(id);
self.state.last_successful_published_time = Some(ele.0.published);
self.state.fail_count = 0;
}
});
Ok(())
}
// this function will return successfully when (a) send succeeded or (b) worker cancelled
// and will return an error if an internal error occurred (send errors cause an infinite loop)
async fn send_retry_loop(
&mut self,
pool: &mut DbPool<'_>,
activity: &SentActivity,
object: &SharedInboxActivities,
inbox_urls: Vec<Url>,
report: UnboundedSender<SendActivityResult>,
initial_fail_count: i32,
domain: String,
context: Data<LemmyContext>,
stop: CancellationToken,
) -> Result<()> {
let inbox_urls = self
.get_inbox_urls(pool, activity)
.await
.context("failed figuring out inbox urls")?;
if inbox_urls.is_empty() {
tracing::debug!("{}: {:?} no inboxes", self.instance.domain, activity.id);
self.state.last_successful_id = Some(activity.id);
self.state.last_successful_published_time = Some(activity.published);
return Ok(());
}
let pool = &mut context.pool();
let Some(actor_apub_id) = &activity.actor_apub_id else {
return Ok(()); // activity was inserted before persistent queue was activated
};
@ -224,61 +391,50 @@ impl InstanceWorker {
.context("failed getting actor instance (was it marked deleted / removed?)")?;
let object = WithContext::new(object.clone(), FEDERATION_CONTEXT.deref().clone());
let inbox_urls = inbox_urls.into_iter().collect();
let requests =
SendActivityTask::prepare(&object, actor.as_ref(), inbox_urls, &self.context).await?;
let requests = SendActivityTask::prepare(&object, actor.as_ref(), inbox_urls, &context).await?;
for task in requests {
// usually only one due to shared inbox
tracing::debug!("sending out {}", task);
while let Err(e) = task.sign_and_send(&self.context).await {
self.state.fail_count += 1;
self.state.last_retry = Some(Utc::now());
let retry_delay: Duration = federate_retry_sleep_duration(self.state.fail_count);
let mut fail_count = initial_fail_count;
while let Err(e) = task.sign_and_send(&context).await {
fail_count += 1;
report.send(SendActivityResult::Failure {
fail_count,
// activity_id: activity.id,
})?;
let retry_delay: Duration = federate_retry_sleep_duration(fail_count);
tracing::info!(
"{}: retrying {:?} attempt {} with delay {retry_delay:.2?}. ({e})",
self.instance.domain,
domain,
activity.id,
self.state.fail_count
fail_count
);
self.save_and_send_state(pool).await?;
tokio::select! {
() = sleep(retry_delay) => {},
() = self.stop.cancelled() => {
() = stop.cancelled() => {
// save state to db and exit
return Ok(());
}
}
}
// Activity send successful, mark instance as alive if it hasn't been updated in a while.
let updated = self.instance.updated.unwrap_or(self.instance.published);
if updated.add(Days::new(1)) < Utc::now() {
self.instance.updated = Some(Utc::now());
let form = InstanceForm::builder()
.domain(self.instance.domain.clone())
.updated(Some(naive_now()))
.build();
Instance::update(pool, self.instance.id, form).await?;
}
}
report.send(SendActivityResult::Success(SendSuccessInfo {
activity_id: activity.id,
published: Some(activity.published),
was_skipped: false,
}))?;
Ok(())
}
/// get inbox urls of sending the given activity to the given instance
/// most often this will return 0 values (if instance doesn't care about the activity)
/// or 1 value (the shared inbox)
/// > 1 values only happens for non-lemmy software
async fn get_inbox_urls(
&mut self,
pool: &mut DbPool<'_>,
activity: &SentActivity,
) -> Result<HashSet<Url>> {
async fn get_inbox_urls(&mut self, activity: &SentActivity) -> Result<HashSet<Url>> {
let mut inbox_urls: HashSet<Url> = HashSet::new();
if activity.send_all_instances {
if !self.site_loaded {
self.site = Site::read_from_instance_id(pool, self.instance.id).await?;
self.site = Site::read_from_instance_id(&mut self.pool(), self.instance.id).await?;
self.site_loaded = true;
}
if let Some(site) = &self.site {
@ -302,23 +458,30 @@ impl InstanceWorker {
Ok(inbox_urls)
}
async fn update_communities(&mut self, pool: &mut DbPool<'_>) -> Result<()> {
async fn update_communities(&mut self) -> Result<()> {
if (Utc::now() - self.last_full_communities_fetch) > *FOLLOW_REMOVALS_RECHECK_DELAY {
tracing::debug!(
"{}: fetching full list of communities",
self.instance.domain
);
// process removals every hour
(self.followed_communities, self.last_full_communities_fetch) = self
.get_communities(pool, self.instance.id, Utc.timestamp_nanos(0))
.get_communities(self.instance.id, Utc.timestamp_nanos(0))
.await?;
self.last_incremental_communities_fetch = self.last_full_communities_fetch;
}
if (Utc::now() - self.last_incremental_communities_fetch) > *FOLLOW_ADDITIONS_RECHECK_DELAY {
// process additions every minute
let (news, time) = self
.get_communities(
pool,
self.instance.id,
self.last_incremental_communities_fetch,
)
.get_communities(self.instance.id, self.last_incremental_communities_fetch)
.await?;
if !news.is_empty() {
tracing::debug!(
"{}: fetched {} incremental new followed communities",
self.instance.domain,
news.len()
);
}
self.followed_communities.extend(news);
self.last_incremental_communities_fetch = time;
}
@ -328,29 +491,38 @@ impl InstanceWorker {
/// get a list of local communities with the remote inboxes on the given instance that cares about them
async fn get_communities(
&mut self,
pool: &mut DbPool<'_>,
instance_id: InstanceId,
last_fetch: DateTime<Utc>,
) -> Result<(HashMap<CommunityId, HashSet<Url>>, DateTime<Utc>)> {
let new_last_fetch =
Utc::now() - chrono::TimeDelta::try_seconds(10).expect("TimeDelta out of bounds"); // update to time before fetch to ensure overlap. subtract 10s to ensure overlap even if published date is not exact
Ok((
CommunityFollowerView::get_instance_followed_community_inboxes(pool, instance_id, last_fetch)
.await?
.into_iter()
.fold(HashMap::new(), |mut map, (c, u)| {
map.entry(c).or_default().insert(u.into());
map
}),
CommunityFollowerView::get_instance_followed_community_inboxes(
&mut self.pool(),
instance_id,
last_fetch,
)
.await?
.into_iter()
.fold(HashMap::new(), |mut map, (c, u)| {
map.entry(c).or_default().insert(u.into());
map
}),
new_last_fetch,
))
}
async fn save_and_send_state(&mut self, pool: &mut DbPool<'_>) -> Result<()> {
async fn save_and_send_state(&mut self) -> Result<()> {
tracing::debug!("{}: saving and sending state", self.instance.domain);
self.last_state_insert = Utc::now();
FederationQueueState::upsert(pool, &self.state).await?;
FederationQueueState::upsert(&mut self.pool(), &self.state).await?;
self
.stats_sender
.send((self.instance.domain.clone(), self.state.clone()))?;
Ok(())
}
fn pool(&self) -> DbPool<'_> {
//self.config.to_request_data()
DbPool::Pool(&self.pool)
}
}