aerogramme/src/bayou.rs

486 lines
17 KiB
Rust
Raw Normal View History

2022-07-13 10:30:35 +00:00
use std::sync::{Arc, Weak};
use std::time::{Duration, Instant};
2022-05-18 10:24:37 +00:00
use anyhow::{anyhow, bail, Result};
2022-07-13 10:30:35 +00:00
use log::{debug, error, info};
2022-05-18 10:24:37 +00:00
use rand::prelude::*;
use serde::{Deserialize, Serialize};
2022-07-13 10:30:35 +00:00
use tokio::sync::{watch, Notify};
2022-05-18 10:24:37 +00:00
use crate::cryptoblob::*;
2022-05-19 11:54:38 +00:00
use crate::login::Credentials;
use crate::timestamp::*;
use crate::storage;
2022-05-18 10:24:37 +00:00
const KEEP_STATE_EVERY: usize = 64;
// Checkpointing interval constants: a checkpoint is not made earlier
// than CHECKPOINT_INTERVAL time after the last one, and is not made
// if there are less than CHECKPOINT_MIN_OPS new operations since last one.
const CHECKPOINT_INTERVAL: Duration = Duration::from_secs(6 * 3600);
2022-05-20 11:36:45 +00:00
const CHECKPOINT_MIN_OPS: usize = 16;
2022-05-18 13:53:13 +00:00
// HYPOTHESIS: processes are able to communicate in a synchronous
// fashion in times that are small compared to CHECKPOINT_INTERVAL.
// More precisely, if a process tried to save an operation within the last
// CHECKPOINT_INTERVAL, we are sure to read it from storage if it was
// successfully saved (and if we don't read it, it means it has been
// definitely discarded due to an error).
// Keep at least two checkpoints, here three, to avoid race conditions
// between processes doing .checkpoint() and those doing .sync()
const CHECKPOINTS_TO_KEEP: usize = 3;
2022-07-13 10:30:35 +00:00
const WATCH_SK: &str = "watch";
2022-05-18 10:24:37 +00:00
pub trait BayouState:
Default + Clone + Serialize + for<'de> Deserialize<'de> + Send + Sync + 'static
{
type Op: Clone + Serialize + for<'de> Deserialize<'de> + std::fmt::Debug + Send + Sync + 'static;
2022-05-18 10:24:37 +00:00
fn apply(&self, op: &Self::Op) -> Self;
}
pub struct Bayou<S: BayouState> {
path: String,
key: Key,
2023-12-18 16:09:44 +00:00
storage: storage::Store,
2022-05-18 10:24:37 +00:00
checkpoint: (Timestamp, S),
history: Vec<(Timestamp, S::Op, Option<S>)>,
2022-07-13 10:30:35 +00:00
last_sync: Option<Instant>,
2022-05-18 14:03:27 +00:00
last_try_checkpoint: Option<Instant>,
2022-07-13 10:30:35 +00:00
watch: Arc<K2vWatch>,
last_sync_watch_ct: storage::RowRef,
2022-05-18 10:24:37 +00:00
}
impl<S: BayouState> Bayou<S> {
2023-12-21 20:54:36 +00:00
pub async fn new(creds: &Credentials, path: String) -> Result<Self> {
let storage = creds.storage.build().await?;
2022-05-18 10:24:37 +00:00
2023-12-18 16:09:44 +00:00
//let target = k2v_client.row(&path, WATCH_SK);
let target = storage::RowRef::new(&path, WATCH_SK);
2023-12-21 20:54:36 +00:00
let watch = K2vWatch::new(creds, target.clone()).await?;
2022-07-13 10:30:35 +00:00
2022-05-18 10:24:37 +00:00
Ok(Self {
path,
2023-12-18 16:09:44 +00:00
storage,
2022-05-19 12:33:49 +00:00
key: creds.keys.master.clone(),
2022-05-18 10:24:37 +00:00
checkpoint: (Timestamp::zero(), S::default()),
history: vec![],
last_sync: None,
2022-05-18 14:03:27 +00:00
last_try_checkpoint: None,
2022-07-13 10:30:35 +00:00
watch,
last_sync_watch_ct: target,
2022-05-18 10:24:37 +00:00
})
}
2022-05-18 10:42:25 +00:00
/// Re-reads the state from persistent storage backend
2022-05-18 10:24:37 +00:00
pub async fn sync(&mut self) -> Result<()> {
let new_last_sync = Some(Instant::now());
let new_last_sync_watch_ct = self.watch.rx.borrow().clone();
2022-05-18 10:24:37 +00:00
// 1. List checkpoints
2022-05-18 13:53:13 +00:00
let checkpoints = self.list_checkpoints().await?;
2022-05-23 16:19:33 +00:00
debug!("(sync) listed checkpoints: {:?}", checkpoints);
2022-05-18 10:42:25 +00:00
// 2. Load last checkpoint if different from currently used one
let checkpoint = if let Some((ts, key)) = checkpoints.last() {
if *ts == self.checkpoint.0 {
(*ts, None)
} else {
2022-05-23 16:19:33 +00:00
debug!("(sync) loading checkpoint: {}", key);
2023-12-18 16:09:44 +00:00
let buf = self.storage.blob_fetch(&storage::BlobRef(key.to_string())).await?.value;
2022-05-23 16:19:33 +00:00
debug!("(sync) checkpoint body length: {}", buf.len());
2022-05-18 14:03:27 +00:00
let ck = open_deserialize::<S>(&buf, &self.key)?;
(*ts, Some(ck))
}
} else {
(Timestamp::zero(), None)
};
if self.checkpoint.0 > checkpoint.0 {
2022-07-13 10:30:35 +00:00
bail!("Loaded checkpoint is more recent than stored one");
}
if let Some(ck) = checkpoint.1 {
2022-05-23 16:19:33 +00:00
debug!(
"(sync) updating checkpoint to loaded state at {:?}",
checkpoint.0
);
self.checkpoint = (checkpoint.0, ck);
};
// remove from history events before checkpoint
self.history = std::mem::take(&mut self.history)
.into_iter()
.skip_while(|(ts, _, _)| *ts < self.checkpoint.0)
.collect();
2022-05-18 10:24:37 +00:00
// 3. List all operations starting from checkpoint
let ts_ser = self.checkpoint.0.to_string();
2022-05-23 16:19:33 +00:00
debug!("(sync) looking up operations starting at {}", ts_ser);
2023-12-18 16:09:44 +00:00
let ops_map = self.storage.row_fetch(&storage::Selector::Range {
shard: &self.path,
sort_begin: &ts_ser,
sort_end: WATCH_SK
}).await?;
let mut ops = vec![];
for row_value in ops_map {
2023-12-18 16:09:44 +00:00
let row = row_value.row_ref;
let sort_key = row.uid.sort;
let ts = sort_key.parse::<Timestamp>().map_err(|_| anyhow!("Invalid operation timestamp: {}", sort_key))?;
2023-12-18 16:09:44 +00:00
let val = row_value.value;
if val.len() != 1 {
2023-12-18 16:09:44 +00:00
bail!("Invalid operation, has {} values", val.len());
}
match &val[0] {
storage::Alternative::Value(v) => {
2023-05-15 16:23:23 +00:00
let op = open_deserialize::<S::Op>(v, &self.key)?;
2023-12-18 16:09:44 +00:00
debug!("(sync) operation {}: {:?}", sort_key, op);
ops.push((ts, op));
}
storage::Alternative::Tombstone => {
continue;
}
}
}
ops.sort_by_key(|(ts, _)| *ts);
2022-05-23 16:19:33 +00:00
debug!("(sync) {} operations", ops.len());
2022-05-18 13:53:13 +00:00
if ops.len() < self.history.len() {
bail!("Some operations have disappeared from storage!");
}
2022-05-18 10:24:37 +00:00
// 4. Check that first operation has same timestamp as checkpoint (if not zero)
if self.checkpoint.0 != Timestamp::zero() && ops[0].0 != self.checkpoint.0 {
bail!(
"First operation in listing doesn't have timestamp that corresponds to checkpoint"
);
}
2022-05-18 10:24:37 +00:00
// 5. Apply all operations in order
// Hypothesis: before the loaded checkpoint, operations haven't changed
// between what's on storage and what we used to calculate the state in RAM here.
let i0 = self
.history
.iter()
.zip(ops.iter())
2022-07-13 10:30:35 +00:00
.take_while(|((ts1, _, _), (ts2, _))| ts1 == ts2)
.count();
if ops.len() > i0 {
// Remove operations from first position where histories differ
self.history.truncate(i0);
// Look up last calculated state which we have saved and start from there.
let mut last_state = (0, &self.checkpoint.1);
for (i, (_, _, state_opt)) in self.history.iter().enumerate().rev() {
if let Some(state) = state_opt {
last_state = (i + 1, state);
break;
}
}
// Calculate state at the end of this common part of the history
let mut state = last_state.1.clone();
for (_, op, _) in self.history[last_state.0..].iter() {
state = state.apply(op);
}
// Now, apply all operations retrieved from storage after the common part
for (ts, op) in ops.drain(i0..) {
state = state.apply(&op);
if (self.history.len() + 1) % KEEP_STATE_EVERY == 0 {
self.history.push((ts, op, Some(state.clone())));
} else {
self.history.push((ts, op, None));
}
}
// Always save final state as result of last operation
self.history.last_mut().unwrap().2 = Some(state);
}
// Save info that sync has been done
self.last_sync = new_last_sync;
2023-12-18 16:09:44 +00:00
self.last_sync_watch_ct = new_last_sync_watch_ct;
Ok(())
}
/// Does a sync() if either of the two conditions is met:
/// - last sync was more than CHECKPOINT_INTERVAL/5 ago
/// - a change was detected
pub async fn opportunistic_sync(&mut self) -> Result<()> {
let too_old = match self.last_sync {
Some(t) => Instant::now() > t + (CHECKPOINT_INTERVAL / 5),
_ => true,
};
2023-12-18 16:09:44 +00:00
let changed = self.last_sync_watch_ct != *self.watch.rx.borrow();
if too_old || changed {
self.sync().await?;
}
Ok(())
2022-05-18 10:24:37 +00:00
}
2022-05-18 10:42:25 +00:00
/// Applies a new operation on the state. Once this function returns,
/// the operation has been safely persisted to storage backend.
/// Make sure to call `.opportunistic_sync()` before doing this,
/// and even before calculating the `op` argument given here.
2022-05-18 10:42:25 +00:00
pub async fn push(&mut self, op: S::Op) -> Result<()> {
2022-05-23 16:19:33 +00:00
debug!("(push) add operation: {:?}", op);
2022-05-18 14:03:27 +00:00
let ts = Timestamp::after(
self.history
.last()
.map(|(ts, _, _)| ts)
.unwrap_or(&self.checkpoint.0),
);
2023-12-18 16:09:44 +00:00
let row_val = storage::RowVal::new(
storage::RowRef::new(&self.path, &ts.to_string()),
seal_serialize(&op, &self.key)?,
);
self.storage.row_insert(vec![row_val]).await?;
2022-07-13 10:30:35 +00:00
self.watch.notify.notify_one();
let new_state = self.state().apply(&op);
self.history.push((ts, op, Some(new_state)));
// Clear previously saved state in history if not required
let hlen = self.history.len();
if hlen >= 2 && (hlen - 1) % KEEP_STATE_EVERY != 0 {
self.history[hlen - 2].2 = None;
}
self.checkpoint().await?;
Ok(())
}
/// Save a new checkpoint if previous checkpoint is too old
pub async fn checkpoint(&mut self) -> Result<()> {
2022-05-18 14:03:27 +00:00
match self.last_try_checkpoint {
Some(ts) if Instant::now() - ts < CHECKPOINT_INTERVAL / 5 => Ok(()),
2022-05-18 14:03:27 +00:00
_ => {
let res = self.checkpoint_internal().await;
if res.is_ok() {
self.last_try_checkpoint = Some(Instant::now());
}
res
}
}
}
async fn checkpoint_internal(&mut self) -> Result<()> {
self.sync().await?;
2022-05-18 13:53:13 +00:00
// Check what would be the possible time for a checkpoint in the history we have
let now = now_msec() as i128;
let i_cp = match self
.history
.iter()
.enumerate()
.rev()
.skip_while(|(_, (ts, _, _))| {
(now - ts.msec as i128) < CHECKPOINT_INTERVAL.as_millis() as i128
})
.map(|(i, _)| i)
.next()
{
Some(i) => i,
None => {
2022-05-23 16:19:33 +00:00
debug!("(cp) Oldest operation is too recent to trigger checkpoint");
2022-05-18 13:53:13 +00:00
return Ok(());
}
};
if i_cp < CHECKPOINT_MIN_OPS {
2022-05-23 16:19:33 +00:00
debug!("(cp) Not enough old operations to trigger checkpoint");
2022-05-18 13:53:13 +00:00
return Ok(());
}
let ts_cp = self.history[i_cp].0;
2022-05-23 16:19:33 +00:00
debug!(
2022-05-18 13:53:13 +00:00
"(cp) we could checkpoint at time {} (index {} in history)",
ts_cp.to_string(),
2022-05-18 13:53:13 +00:00
i_cp
);
// Check existing checkpoints: if last one is too recent, don't checkpoint again.
let existing_checkpoints = self.list_checkpoints().await?;
2022-05-23 16:19:33 +00:00
debug!("(cp) listed checkpoints: {:?}", existing_checkpoints);
2022-05-18 13:53:13 +00:00
if let Some(last_cp) = existing_checkpoints.last() {
if (ts_cp.msec as i128 - last_cp.0.msec as i128)
< CHECKPOINT_INTERVAL.as_millis() as i128
{
2022-05-23 16:19:33 +00:00
debug!(
2022-05-18 13:53:13 +00:00
"(cp) last checkpoint is too recent: {}, not checkpointing",
last_cp.0.to_string()
2022-05-18 13:53:13 +00:00
);
return Ok(());
}
}
debug!("(cp) saving checkpoint at {}", ts_cp.to_string());
2022-05-18 13:53:13 +00:00
// Calculate state at time of checkpoint
let mut last_known_state = (0, &self.checkpoint.1);
for (i, (_, _, st)) in self.history[..i_cp].iter().enumerate() {
if let Some(s) = st {
last_known_state = (i + 1, s);
}
}
let mut state_cp = last_known_state.1.clone();
for (_, op, _) in self.history[last_known_state.0..i_cp].iter() {
state_cp = state_cp.apply(op);
}
// Serialize and save checkpoint
let cryptoblob = seal_serialize(&state_cp, &self.key)?;
2022-05-23 16:19:33 +00:00
debug!("(cp) checkpoint body length: {}", cryptoblob.len());
2022-05-18 13:53:13 +00:00
2023-12-18 16:09:44 +00:00
let blob_val = storage::BlobVal::new(
storage::BlobRef(format!("{}/checkpoint/{}", self.path, ts_cp.to_string())),
cryptoblob.into(),
);
self.storage.blob_insert(&blob_val).await?;
2022-05-18 13:53:13 +00:00
// Drop old checkpoints (but keep at least CHECKPOINTS_TO_KEEP of them)
let ecp_len = existing_checkpoints.len();
if ecp_len + 1 > CHECKPOINTS_TO_KEEP {
let last_to_keep = ecp_len + 1 - CHECKPOINTS_TO_KEEP;
// Delete blobs
for (_ts, key) in existing_checkpoints[..last_to_keep].iter() {
2022-05-23 16:19:33 +00:00
debug!("(cp) drop old checkpoint {}", key);
2023-12-18 16:09:44 +00:00
self.storage.blob_rm(&storage::BlobRef(key.to_string())).await?;
2022-05-18 13:53:13 +00:00
}
// Delete corresponding range of operations
let ts_ser = existing_checkpoints[last_to_keep].0.to_string();
2023-12-18 16:09:44 +00:00
self.storage.row_rm(&storage::Selector::Range {
shard: &self.path,
sort_begin: "",
sort_end: &ts_ser
}).await?
2022-05-18 13:53:13 +00:00
}
Ok(())
2022-05-18 10:24:37 +00:00
}
2022-05-18 10:42:25 +00:00
pub fn state(&self) -> &S {
if let Some(last) = self.history.last() {
last.2.as_ref().unwrap()
} else {
&self.checkpoint.1
}
}
2022-05-18 13:53:13 +00:00
// ---- INTERNAL ----
async fn list_checkpoints(&self) -> Result<Vec<(Timestamp, String)>> {
let prefix = format!("{}/checkpoint/", self.path);
2023-12-18 16:09:44 +00:00
let checkpoints_res = self.storage.blob_list(&prefix).await?;
2022-05-18 13:53:13 +00:00
let mut checkpoints = vec![];
for object in checkpoints_res {
2023-12-18 16:09:44 +00:00
let key = object.0;
if let Some(ckid) = key.strip_prefix(&prefix) {
if let Ok(ts) = ckid.parse::<Timestamp>() {
checkpoints.push((ts, key.into()));
2022-05-18 13:53:13 +00:00
}
}
}
checkpoints.sort_by_key(|(ts, _)| *ts);
Ok(checkpoints)
}
2022-05-18 10:24:37 +00:00
}
2022-07-13 10:30:35 +00:00
// ---- Bayou watch in K2V ----
struct K2vWatch {
2023-12-18 16:09:44 +00:00
target: storage::RowRef,
rx: watch::Receiver<storage::RowRef>,
2022-07-13 10:30:35 +00:00
notify: Notify,
}
impl K2vWatch {
/// Creates a new watch and launches subordinate threads.
/// These threads hold Weak pointers to the struct;
/// they exit when the Arc is dropped.
2023-12-21 20:54:36 +00:00
async fn new(creds: &Credentials, target: storage::RowRef) -> Result<Arc<Self>> {
let storage = creds.storage.build().await?;
2023-12-18 16:09:44 +00:00
let (tx, rx) = watch::channel::<storage::RowRef>(target.clone());
2022-07-13 10:30:35 +00:00
let notify = Notify::new();
2023-12-18 16:09:44 +00:00
let watch = Arc::new(K2vWatch { target, rx, notify });
2022-07-13 10:30:35 +00:00
tokio::spawn(Self::background_task(
2022-07-13 10:30:35 +00:00
Arc::downgrade(&watch),
2023-12-18 16:09:44 +00:00
storage,
2022-07-13 10:30:35 +00:00
tx,
));
Ok(watch)
}
async fn background_task(
2022-07-13 10:30:35 +00:00
self_weak: Weak<Self>,
2023-12-18 16:09:44 +00:00
storage: storage::Store,
tx: watch::Sender<storage::RowRef>,
2022-07-13 10:30:35 +00:00
) {
let mut row = match Weak::upgrade(&self_weak) {
2023-12-18 16:09:44 +00:00
Some(this) => this.target.clone(),
None => {
error!("can't start loop");
return
},
};
2022-07-13 10:30:35 +00:00
while let Some(this) = Weak::upgrade(&self_weak) {
debug!(
"bayou k2v watch bg loop iter ({}, {})",
2023-12-18 16:09:44 +00:00
this.target.uid.shard, this.target.uid.sort
2022-07-13 10:30:35 +00:00
);
tokio::select!(
_ = tokio::time::sleep(Duration::from_secs(60)) => continue,
2023-12-18 16:09:44 +00:00
update = storage.row_poll(&row) => {
match update {
Err(e) => {
error!("Error in bayou k2v wait value changed: {}", e);
tokio::time::sleep(Duration::from_secs(30)).await;
}
Ok(new_value) => {
2023-12-18 16:09:44 +00:00
row = new_value.row_ref;
if tx.send(row.clone()).is_err() {
break;
}
}
2022-07-13 10:30:35 +00:00
}
}
_ = this.notify.notified() => {
let rand = u128::to_be_bytes(thread_rng().gen()).to_vec();
2023-12-18 16:09:44 +00:00
let row_val = storage::RowVal::new(row.clone(), rand);
if let Err(e) = storage.row_insert(vec![row_val]).await
2022-07-13 10:30:35 +00:00
{
error!("Error in bayou k2v watch updater loop: {}", e);
tokio::time::sleep(Duration::from_secs(30)).await;
}
}
);
2022-07-13 10:30:35 +00:00
}
info!("bayou k2v watch bg loop exiting");
2022-07-13 10:30:35 +00:00
}
}