garage/src/table/sync.rs

672 lines
16 KiB
Rust
Raw Normal View History

use std::collections::VecDeque;
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant};
2021-10-14 09:50:12 +00:00
use async_trait::async_trait;
use futures::select;
use futures_util::future::*;
use futures_util::stream::*;
2022-02-16 13:23:04 +00:00
use opentelemetry::KeyValue;
use rand::Rng;
use serde::{Deserialize, Serialize};
use serde_bytes::ByteBuf;
use tokio::sync::{mpsc, watch};
use garage_util::data::*;
use garage_util::error::Error;
2021-03-16 11:18:03 +00:00
use garage_rpc::ring::*;
2021-10-14 09:50:12 +00:00
use garage_rpc::system::System;
use garage_rpc::*;
use crate::data::*;
use crate::merkle::*;
use crate::replication::*;
use crate::*;
const TABLE_SYNC_RPC_TIMEOUT: Duration = Duration::from_secs(30);
// Do anti-entropy every 10 minutes
const ANTI_ENTROPY_INTERVAL: Duration = Duration::from_secs(10 * 60);
2021-10-14 09:50:12 +00:00
pub struct TableSyncer<F: TableSchema + 'static, R: TableReplication + 'static> {
2021-03-16 10:43:58 +00:00
system: Arc<System>,
data: Arc<TableData<F, R>>,
merkle: Arc<MerkleUpdater<F, R>>,
todo: Mutex<SyncTodo>,
2021-10-14 09:50:12 +00:00
endpoint: Arc<Endpoint<SyncRpc, Self>>,
}
#[derive(Serialize, Deserialize)]
2021-05-02 21:13:08 +00:00
pub(crate) enum SyncRpc {
2021-03-16 11:18:03 +00:00
RootCkHash(Partition, Hash),
RootCkDifferent(bool),
GetNode(MerkleNodeKey),
Node(MerkleNodeKey, MerkleNode),
Items(Vec<Arc<ByteBuf>>),
Ok,
}
impl Rpc for SyncRpc {
type Response = Result<SyncRpc, Error>;
2021-10-14 09:50:12 +00:00
}
struct SyncTodo {
todo: Vec<TodoPartition>,
}
#[derive(Debug, Clone)]
struct TodoPartition {
2021-03-16 11:18:03 +00:00
partition: Partition,
begin: Hash,
end: Hash,
// Are we a node that stores this partition or not?
retain: bool,
}
impl<F, R> TableSyncer<F, R>
where
F: TableSchema + 'static,
R: TableReplication + 'static,
{
pub(crate) fn launch(
2021-03-16 10:43:58 +00:00
system: Arc<System>,
data: Arc<TableData<F, R>>,
merkle: Arc<MerkleUpdater<F, R>>,
) -> Arc<Self> {
2021-10-14 09:50:12 +00:00
let endpoint = system
.netapp
2021-12-14 11:34:01 +00:00
.endpoint(format!("garage_table/sync.rs/Rpc:{}", F::TABLE_NAME));
let todo = SyncTodo { todo: vec![] };
let syncer = Arc::new(Self {
2021-03-16 10:43:58 +00:00
system: system.clone(),
2021-12-14 11:34:01 +00:00
data,
2021-03-16 10:43:58 +00:00
merkle,
todo: Mutex::new(todo),
2021-10-14 09:50:12 +00:00
endpoint,
});
2021-10-14 09:50:12 +00:00
syncer.endpoint.set_handler(syncer.clone());
let (busy_tx, busy_rx) = mpsc::unbounded_channel();
let s1 = syncer.clone();
2021-03-16 10:43:58 +00:00
system.background.spawn_worker(
2021-12-14 11:34:01 +00:00
format!("table sync watcher for {}", F::TABLE_NAME),
move |must_exit: watch::Receiver<bool>| s1.watcher_task(must_exit, busy_rx),
);
let s2 = syncer.clone();
2021-03-16 10:43:58 +00:00
system.background.spawn_worker(
2021-12-14 11:34:01 +00:00
format!("table syncer for {}", F::TABLE_NAME),
move |must_exit: watch::Receiver<bool>| s2.syncer_task(must_exit, busy_tx),
);
let s3 = syncer.clone();
tokio::spawn(async move {
2021-03-15 21:36:41 +00:00
tokio::time::sleep(Duration::from_secs(20)).await;
s3.add_full_sync();
});
syncer
}
async fn watcher_task(
self: Arc<Self>,
mut must_exit: watch::Receiver<bool>,
mut busy_rx: mpsc::UnboundedReceiver<bool>,
) {
2021-03-16 10:43:58 +00:00
let mut prev_ring: Arc<Ring> = self.system.ring.borrow().clone();
let mut ring_recv: watch::Receiver<Arc<Ring>> = self.system.ring.clone();
let mut nothing_to_do_since = Some(Instant::now());
while !*must_exit.borrow() {
select! {
2021-03-15 21:36:41 +00:00
_ = ring_recv.changed().fuse() => {
let new_ring = ring_recv.borrow();
if !Arc::ptr_eq(&new_ring, &prev_ring) {
2021-12-14 11:34:01 +00:00
debug!("({}) Ring changed, adding full sync to syncer todo list", F::TABLE_NAME);
2021-03-15 21:36:41 +00:00
self.add_full_sync();
prev_ring = new_ring.clone();
}
}
2021-03-15 21:36:41 +00:00
busy_opt = busy_rx.recv().fuse() => {
if let Some(busy) = busy_opt {
if busy {
nothing_to_do_since = None;
2021-04-23 19:42:52 +00:00
} else if nothing_to_do_since.is_none() {
nothing_to_do_since = Some(Instant::now());
}
}
}
2021-04-23 19:42:52 +00:00
_ = must_exit.changed().fuse() => {},
2021-03-15 21:36:41 +00:00
_ = tokio::time::sleep(Duration::from_secs(1)).fuse() => {
if nothing_to_do_since.map(|t| Instant::now() - t >= ANTI_ENTROPY_INTERVAL).unwrap_or(false) {
nothing_to_do_since = None;
2021-12-14 11:34:01 +00:00
debug!("({}) Interval passed, adding full sync to syncer todo list", F::TABLE_NAME);
self.add_full_sync();
}
}
}
}
}
pub fn add_full_sync(&self) {
self.todo
.lock()
.unwrap()
2021-03-16 10:43:58 +00:00
.add_full_sync(&self.data, &self.system);
}
async fn syncer_task(
self: Arc<Self>,
mut must_exit: watch::Receiver<bool>,
busy_tx: mpsc::UnboundedSender<bool>,
) {
while !*must_exit.borrow() {
let task = self.todo.lock().unwrap().pop_task();
if let Some(partition) = task {
busy_tx.send(true).unwrap();
let res = self
.clone()
.sync_partition(&partition, &mut must_exit)
.await;
if let Err(e) = res {
warn!(
"({}) Error while syncing {:?}: {}",
2021-12-14 11:34:01 +00:00
F::TABLE_NAME,
partition,
e
);
}
} else {
busy_tx.send(false).unwrap();
2021-03-15 21:36:41 +00:00
tokio::time::sleep(Duration::from_secs(1)).await;
}
}
}
async fn sync_partition(
self: Arc<Self>,
partition: &TodoPartition,
must_exit: &mut watch::Receiver<bool>,
) -> Result<(), Error> {
if partition.retain {
2021-03-16 10:43:58 +00:00
let my_id = self.system.id;
let nodes = self
2021-03-16 10:43:58 +00:00
.data
.replication
2021-03-16 11:18:03 +00:00
.write_nodes(&partition.begin)
.into_iter()
.filter(|node| *node != my_id)
.collect::<Vec<_>>();
debug!(
"({}) Syncing {:?} with {:?}...",
2021-12-14 11:34:01 +00:00
F::TABLE_NAME,
partition,
nodes
);
let mut sync_futures = nodes
.iter()
.map(|node| {
self.clone()
.do_sync_with(partition.clone(), *node, must_exit.clone())
})
.collect::<FuturesUnordered<_>>();
let mut n_errors = 0;
while let Some(r) = sync_futures.next().await {
if let Err(e) = r {
n_errors += 1;
2021-12-14 11:34:01 +00:00
warn!("({}) Sync error: {}", F::TABLE_NAME, e);
}
}
2021-03-16 10:43:58 +00:00
if n_errors > self.data.replication.max_write_errors() {
return Err(Error::Message(format!(
"Sync failed with too many nodes (should have been: {:?}).",
nodes
)));
}
} else {
2021-03-16 14:58:40 +00:00
self.offload_partition(&partition.begin, &partition.end, must_exit)
.await?;
}
Ok(())
}
// Offload partition: this partition is not something we are storing,
// so send it out to all other nodes that store it and delete items locally.
// We don't bother checking if the remote nodes already have the items,
// we just batch-send everything. Offloading isn't supposed to happen very often.
// If any of the nodes that are supposed to store the items is unable to
// save them, we interrupt the process.
async fn offload_partition(
self: &Arc<Self>,
begin: &Hash,
end: &Hash,
must_exit: &mut watch::Receiver<bool>,
) -> Result<(), Error> {
let mut counter: usize = 0;
while !*must_exit.borrow() {
let mut items = Vec::new();
Abstract database behind generic interface and implement alternative drivers (#322) - [x] Design interface - [x] Implement Sled backend - [x] Re-implement the SledCountedTree hack ~~on Sled backend~~ on all backends (i.e. over the abstraction) - [x] Convert Garage code to use generic interface - [x] Proof-read converted Garage code - [ ] Test everything well - [x] Implement sqlite backend - [x] Implement LMDB backend - [ ] (Implement Persy backend?) - [ ] (Implement other backends? (like RocksDB, ...)) - [x] Implement backend choice in config file and garage server module - [x] Add CLI for converting between DB formats - Exploit the new interface to put more things in transactions - [x] `.updated()` trigger on Garage tables Fix #284 **Bugs** - [x] When exporting sqlite, trees iterate empty?? - [x] LMDB doesn't work **Known issues for various back-ends** - Sled: - Eats all my RAM and also all my disk space - `.len()` has to traverse the whole table - Is actually quite slow on some operations - And is actually pretty bad code... - Sqlite: - Requires a lock to be taken on all operations. The lock is also taken when iterating on a table with `.iter()`, and the lock isn't released until the iterator is dropped. This means that we must be VERY carefull to not do anything else inside a `.iter()` loop or else we will have a deadlock! Most such cases have been eliminated from the Garage codebase, but there might still be some that remain. If your Garage-over-Sqlite seems to hang/freeze, this is the reason. - (adapter uses a bunch of unsafe code) - Heed (LMDB): - Not suited for 32-bit machines as it has to map the whole DB in memory. - (adpater uses a tiny bit of unsafe code) **My recommendation:** avoid 32-bit machines and use LMDB as much as possible. **Converting databases** is actually quite easy. For example from Sled to LMDB: ```bash cd src/db cargo run --features cli --bin convert -- -i path/to/garage/meta/db -a sled -o path/to/garage/meta/db.lmdb -b lmdb ``` Then, just add this to your `config.toml`: ```toml db_engine = "lmdb" ``` Co-authored-by: Alex Auvolat <alex@adnab.me> Reviewed-on: https://git.deuxfleurs.fr/Deuxfleurs/garage/pulls/322 Co-authored-by: Alex <alex@adnab.me> Co-committed-by: Alex <alex@adnab.me>
2022-06-08 08:01:44 +00:00
for item in self.data.store.range(begin.to_vec()..end.to_vec())? {
let (key, value) = item?;
Abstract database behind generic interface and implement alternative drivers (#322) - [x] Design interface - [x] Implement Sled backend - [x] Re-implement the SledCountedTree hack ~~on Sled backend~~ on all backends (i.e. over the abstraction) - [x] Convert Garage code to use generic interface - [x] Proof-read converted Garage code - [ ] Test everything well - [x] Implement sqlite backend - [x] Implement LMDB backend - [ ] (Implement Persy backend?) - [ ] (Implement other backends? (like RocksDB, ...)) - [x] Implement backend choice in config file and garage server module - [x] Add CLI for converting between DB formats - Exploit the new interface to put more things in transactions - [x] `.updated()` trigger on Garage tables Fix #284 **Bugs** - [x] When exporting sqlite, trees iterate empty?? - [x] LMDB doesn't work **Known issues for various back-ends** - Sled: - Eats all my RAM and also all my disk space - `.len()` has to traverse the whole table - Is actually quite slow on some operations - And is actually pretty bad code... - Sqlite: - Requires a lock to be taken on all operations. The lock is also taken when iterating on a table with `.iter()`, and the lock isn't released until the iterator is dropped. This means that we must be VERY carefull to not do anything else inside a `.iter()` loop or else we will have a deadlock! Most such cases have been eliminated from the Garage codebase, but there might still be some that remain. If your Garage-over-Sqlite seems to hang/freeze, this is the reason. - (adapter uses a bunch of unsafe code) - Heed (LMDB): - Not suited for 32-bit machines as it has to map the whole DB in memory. - (adpater uses a tiny bit of unsafe code) **My recommendation:** avoid 32-bit machines and use LMDB as much as possible. **Converting databases** is actually quite easy. For example from Sled to LMDB: ```bash cd src/db cargo run --features cli --bin convert -- -i path/to/garage/meta/db -a sled -o path/to/garage/meta/db.lmdb -b lmdb ``` Then, just add this to your `config.toml`: ```toml db_engine = "lmdb" ``` Co-authored-by: Alex Auvolat <alex@adnab.me> Reviewed-on: https://git.deuxfleurs.fr/Deuxfleurs/garage/pulls/322 Co-authored-by: Alex <alex@adnab.me> Co-committed-by: Alex <alex@adnab.me>
2022-06-08 08:01:44 +00:00
items.push((key.to_vec(), Arc::new(ByteBuf::from(value))));
if items.len() >= 1024 {
break;
}
}
2021-04-23 19:42:52 +00:00
if !items.is_empty() {
let nodes = self
2021-03-16 10:43:58 +00:00
.data
.replication
2021-10-26 08:20:05 +00:00
.write_nodes(begin)
.into_iter()
.collect::<Vec<_>>();
2021-03-16 10:43:58 +00:00
if nodes.contains(&self.system.id) {
warn!(
"({}) Interrupting offload as partitions seem to have changed",
2021-12-14 11:34:01 +00:00
F::TABLE_NAME
);
break;
}
2021-03-16 10:43:58 +00:00
if nodes.len() < self.data.replication.write_quorum() {
2021-04-23 19:42:52 +00:00
return Err(Error::Message(
"Not offloading as we don't have a quorum of nodes to write to."
2021-04-23 19:42:52 +00:00
.to_string(),
));
}
counter += 1;
info!(
"({}) Offloading {} items from {:?}..{:?} ({})",
2021-12-14 11:34:01 +00:00
F::TABLE_NAME,
items.len(),
begin,
end,
counter
);
self.offload_items(&items, &nodes[..]).await?;
} else {
break;
}
}
Ok(())
}
async fn offload_items(
self: &Arc<Self>,
2021-04-23 19:42:52 +00:00
items: &[(Vec<u8>, Arc<ByteBuf>)],
nodes: &[Uuid],
) -> Result<(), Error> {
let values = items.iter().map(|(_k, v)| v.clone()).collect::<Vec<_>>();
2022-02-16 13:23:04 +00:00
for to in nodes.iter() {
self.data.metrics.sync_items_sent.add(
values.len() as u64,
&[
KeyValue::new("table_name", F::TABLE_NAME),
KeyValue::new("to", format!("{:?}", to)),
],
);
}
2021-10-14 09:50:12 +00:00
self.system
.rpc
2021-03-12 20:52:19 +00:00
.try_call_many(
2021-10-14 09:50:12 +00:00
&self.endpoint,
2021-04-23 19:42:52 +00:00
nodes,
2021-05-02 21:13:08 +00:00
SyncRpc::Items(values),
2021-10-14 09:50:12 +00:00
RequestStrategy::with_priority(PRIO_BACKGROUND)
.with_quorum(nodes.len())
.with_timeout(TABLE_SYNC_RPC_TIMEOUT),
2021-03-12 20:52:19 +00:00
)
.await?;
// All remote nodes have written those items, now we can delete them locally
let mut not_removed = 0;
for (k, v) in items.iter() {
if !self.data.delete_if_equal(&k[..], &v[..])? {
not_removed += 1;
}
}
if not_removed > 0 {
2021-12-14 11:34:01 +00:00
debug!("({}) {} items not removed during offload because they changed in between (trying again...)", F::TABLE_NAME, not_removed);
}
Ok(())
}
// ======= SYNCHRONIZATION PROCEDURE -- DRIVER SIDE ======
2021-03-11 17:45:26 +00:00
// The driver side is only concerned with sending out the item it has
// and the other side might not have. Receiving items that differ from one
// side to the other will happen when the other side syncs with us,
// which they also do regularly.
2021-03-16 11:18:03 +00:00
fn get_root_ck(&self, partition: Partition) -> Result<(MerkleNodeKey, MerkleNode), Error> {
let key = MerkleNodeKey {
partition,
prefix: vec![],
};
2021-03-16 11:18:03 +00:00
let node = self.merkle.read_node(&key)?;
Ok((key, node))
}
async fn do_sync_with(
self: Arc<Self>,
partition: TodoPartition,
who: Uuid,
must_exit: watch::Receiver<bool>,
) -> Result<(), Error> {
2021-03-16 11:18:03 +00:00
let (root_ck_key, root_ck) = self.get_root_ck(partition.partition)?;
2021-03-11 18:30:24 +00:00
if root_ck.is_empty() {
debug!(
"({}) Sync {:?} with {:?}: partition is empty.",
2021-12-14 11:34:01 +00:00
F::TABLE_NAME,
partition,
who
2021-03-11 18:30:24 +00:00
);
return Ok(());
2021-03-11 18:30:24 +00:00
}
2021-03-16 11:18:03 +00:00
let root_ck_hash = hash_of::<MerkleNode>(&root_ck)?;
2021-03-11 18:30:24 +00:00
2021-03-16 11:18:03 +00:00
// Check if they have the same root checksum
// If so, do nothing.
let root_resp = self
2021-10-14 09:50:12 +00:00
.system
.rpc
.call(
2021-10-14 09:50:12 +00:00
&self.endpoint,
who,
2021-05-02 21:13:08 +00:00
SyncRpc::RootCkHash(partition.partition, root_ck_hash),
2021-10-14 09:50:12 +00:00
RequestStrategy::with_priority(PRIO_BACKGROUND)
.with_timeout(TABLE_SYNC_RPC_TIMEOUT),
)
.await?;
let mut todo = match root_resp {
2021-05-02 21:13:08 +00:00
SyncRpc::RootCkDifferent(false) => {
debug!(
"({}) Sync {:?} with {:?}: no difference",
2021-12-14 11:34:01 +00:00
F::TABLE_NAME,
partition,
who
);
return Ok(());
}
2021-05-02 21:13:08 +00:00
SyncRpc::RootCkDifferent(true) => VecDeque::from(vec![root_ck_key]),
x => {
return Err(Error::Message(format!(
"Invalid respone to RootCkHash RPC: {}",
debug_serialize(x)
)));
}
};
let mut todo_items = vec![];
while !todo.is_empty() && !*must_exit.borrow() {
let key = todo.pop_front().unwrap();
2021-03-16 10:43:58 +00:00
let node = self.merkle.read_node(&key)?;
match node {
MerkleNode::Empty => {
// They have items we don't have.
// We don't request those items from them, they will send them.
// We only bother with pushing items that differ
}
2021-03-11 17:50:32 +00:00
MerkleNode::Leaf(ik, ivhash) => {
// Just send that item directly
2021-03-11 17:50:32 +00:00
if let Some(val) = self.data.store.get(&ik[..])? {
if blake2sum(&val[..]) != ivhash {
2021-12-14 11:34:01 +00:00
warn!("({}) Hashes differ between stored value and Merkle tree, key: {:?} (if your server is very busy, don't worry, this happens when the Merkle tree can't be updated fast enough)", F::TABLE_NAME, ik);
2021-03-11 17:50:32 +00:00
}
todo_items.push(val.to_vec());
} else {
2021-12-14 11:34:01 +00:00
warn!("({}) Item from Merkle tree not found in store: {:?} (if your server is very busy, don't worry, this happens when the Merkle tree can't be updated fast enough)", F::TABLE_NAME, ik);
}
}
MerkleNode::Intermediate(l) => {
// Get Merkle node for this tree position at remote node
// and compare it with local node
let remote_node = match self
2021-10-14 09:50:12 +00:00
.system
.rpc
.call(
&self.endpoint,
who,
SyncRpc::GetNode(key.clone()),
RequestStrategy::with_priority(PRIO_BACKGROUND)
.with_timeout(TABLE_SYNC_RPC_TIMEOUT),
)
.await?
{
2021-05-02 21:13:08 +00:00
SyncRpc::Node(_, node) => node,
x => {
return Err(Error::Message(format!(
"Invalid respone to GetNode RPC: {}",
debug_serialize(x)
)));
}
};
let int_l2 = match remote_node {
// If they have an intermediate node at this tree position,
// we can compare them to find differences
MerkleNode::Intermediate(l2) => l2,
// Otherwise, treat it as if they have nothing for this subtree,
// which will have the consequence of sending them everything
_ => vec![],
};
let join = join_ordered(&l[..], &int_l2[..]);
for (p, v1, v2) in join.into_iter() {
let diff = match (v1, v2) {
(Some(_), None) | (None, Some(_)) => true,
(Some(a), Some(b)) => a != b,
_ => false,
};
if diff {
todo.push_back(key.add_byte(*p));
}
}
}
}
if todo_items.len() >= 256 {
2021-04-23 19:42:52 +00:00
self.send_items(who, std::mem::take(&mut todo_items))
.await?;
}
}
if !todo_items.is_empty() {
self.send_items(who, todo_items).await?;
}
Ok(())
}
async fn send_items(&self, who: Uuid, item_value_list: Vec<Vec<u8>>) -> Result<(), Error> {
info!(
"({}) Sending {} items to {:?}",
2021-12-14 11:34:01 +00:00
F::TABLE_NAME,
item_value_list.len(),
who
);
let values = item_value_list
.into_iter()
.map(|x| Arc::new(ByteBuf::from(x)))
.collect::<Vec<_>>();
2022-02-16 13:23:04 +00:00
self.data.metrics.sync_items_sent.add(
values.len() as u64,
&[
KeyValue::new("table_name", F::TABLE_NAME),
KeyValue::new("to", format!("{:?}", who)),
],
);
let rpc_resp = self
2021-10-14 09:50:12 +00:00
.system
.rpc
.call(
&self.endpoint,
who,
SyncRpc::Items(values),
RequestStrategy::with_priority(PRIO_BACKGROUND)
.with_timeout(TABLE_SYNC_RPC_TIMEOUT),
)
.await?;
2021-05-02 21:13:08 +00:00
if let SyncRpc::Ok = rpc_resp {
Ok(())
} else {
Err(Error::unexpected_rpc_message(rpc_resp))
}
}
}
// ======= SYNCHRONIZATION PROCEDURE -- RECEIVER SIDE ======
#[async_trait]
impl<F, R> EndpointHandler<SyncRpc> for TableSyncer<F, R>
where
F: TableSchema + 'static,
R: TableReplication + 'static,
{
2022-02-16 13:23:04 +00:00
async fn handle(self: &Arc<Self>, message: &SyncRpc, from: NodeID) -> Result<SyncRpc, Error> {
match message {
2021-05-02 21:13:08 +00:00
SyncRpc::RootCkHash(range, h) => {
2021-03-16 11:18:03 +00:00
let (_root_ck_key, root_ck) = self.get_root_ck(*range)?;
let hash = hash_of::<MerkleNode>(&root_ck)?;
2021-05-02 21:13:08 +00:00
Ok(SyncRpc::RootCkDifferent(hash != *h))
}
2021-05-02 21:13:08 +00:00
SyncRpc::GetNode(k) => {
2021-10-26 08:20:05 +00:00
let node = self.merkle.read_node(k)?;
2021-05-02 21:13:08 +00:00
Ok(SyncRpc::Node(k.clone(), node))
}
2021-05-02 21:13:08 +00:00
SyncRpc::Items(items) => {
2022-02-16 13:23:04 +00:00
self.data.metrics.sync_items_received.add(
items.len() as u64,
&[
KeyValue::new("table_name", F::TABLE_NAME),
KeyValue::new(
"from",
format!("{:?}", Uuid::try_from(from.as_ref()).unwrap()),
),
],
);
self.data.update_many(items)?;
2021-05-02 21:13:08 +00:00
Ok(SyncRpc::Ok)
}
m => Err(Error::unexpected_rpc_message(m)),
}
}
}
impl SyncTodo {
fn add_full_sync<F: TableSchema, R: TableReplication>(
&mut self,
2021-03-16 10:43:58 +00:00
data: &TableData<F, R>,
system: &System,
) {
2021-03-16 10:43:58 +00:00
let my_id = system.id;
self.todo.clear();
2021-03-16 11:18:03 +00:00
let partitions = data.replication.partitions();
2021-03-16 11:18:03 +00:00
for i in 0..partitions.len() {
let begin = partitions[i].1;
2021-03-16 11:18:03 +00:00
let end = if i + 1 < partitions.len() {
2021-03-16 14:58:40 +00:00
partitions[i + 1].1
} else {
2021-03-16 11:18:03 +00:00
[0xFFu8; 32].into()
};
2021-03-16 11:18:03 +00:00
let nodes = data.replication.write_nodes(&begin);
let retain = nodes.contains(&my_id);
if !retain {
// Check if we have some data to send, otherwise skip
Abstract database behind generic interface and implement alternative drivers (#322) - [x] Design interface - [x] Implement Sled backend - [x] Re-implement the SledCountedTree hack ~~on Sled backend~~ on all backends (i.e. over the abstraction) - [x] Convert Garage code to use generic interface - [x] Proof-read converted Garage code - [ ] Test everything well - [x] Implement sqlite backend - [x] Implement LMDB backend - [ ] (Implement Persy backend?) - [ ] (Implement other backends? (like RocksDB, ...)) - [x] Implement backend choice in config file and garage server module - [x] Add CLI for converting between DB formats - Exploit the new interface to put more things in transactions - [x] `.updated()` trigger on Garage tables Fix #284 **Bugs** - [x] When exporting sqlite, trees iterate empty?? - [x] LMDB doesn't work **Known issues for various back-ends** - Sled: - Eats all my RAM and also all my disk space - `.len()` has to traverse the whole table - Is actually quite slow on some operations - And is actually pretty bad code... - Sqlite: - Requires a lock to be taken on all operations. The lock is also taken when iterating on a table with `.iter()`, and the lock isn't released until the iterator is dropped. This means that we must be VERY carefull to not do anything else inside a `.iter()` loop or else we will have a deadlock! Most such cases have been eliminated from the Garage codebase, but there might still be some that remain. If your Garage-over-Sqlite seems to hang/freeze, this is the reason. - (adapter uses a bunch of unsafe code) - Heed (LMDB): - Not suited for 32-bit machines as it has to map the whole DB in memory. - (adpater uses a tiny bit of unsafe code) **My recommendation:** avoid 32-bit machines and use LMDB as much as possible. **Converting databases** is actually quite easy. For example from Sled to LMDB: ```bash cd src/db cargo run --features cli --bin convert -- -i path/to/garage/meta/db -a sled -o path/to/garage/meta/db.lmdb -b lmdb ``` Then, just add this to your `config.toml`: ```toml db_engine = "lmdb" ``` Co-authored-by: Alex Auvolat <alex@adnab.me> Reviewed-on: https://git.deuxfleurs.fr/Deuxfleurs/garage/pulls/322 Co-authored-by: Alex <alex@adnab.me> Co-committed-by: Alex <alex@adnab.me>
2022-06-08 08:01:44 +00:00
match data.store.range(begin..end) {
Ok(mut iter) => {
if iter.next().is_none() {
continue;
}
}
Err(e) => {
warn!("DB error in add_full_sync: {}", e);
continue;
}
}
}
self.todo.push(TodoPartition {
2021-03-16 11:18:03 +00:00
partition: partitions[i].0,
begin,
end,
retain,
});
}
}
fn pop_task(&mut self) -> Option<TodoPartition> {
if self.todo.is_empty() {
return None;
}
2021-03-16 14:58:40 +00:00
let i = rand::thread_rng().gen_range(0..self.todo.len());
if i == self.todo.len() - 1 {
self.todo.pop()
} else {
let replacement = self.todo.pop().unwrap();
let ret = std::mem::replace(&mut self.todo[i], replacement);
Some(ret)
}
}
}
fn hash_of<T: Serialize>(x: &T) -> Result<Hash, Error> {
Ok(blake2sum(&rmp_to_vec_all_named(x)?[..]))
}
fn join_ordered<'a, K: Ord + Eq, V1, V2>(
x: &'a [(K, V1)],
y: &'a [(K, V2)],
) -> Vec<(&'a K, Option<&'a V1>, Option<&'a V2>)> {
let mut ret = vec![];
let mut i = 0;
let mut j = 0;
while i < x.len() || j < y.len() {
if i < x.len() && j < y.len() && x[i].0 == y[j].0 {
ret.push((&x[i].0, Some(&x[i].1), Some(&y[j].1)));
i += 1;
j += 1;
} else if i < x.len() && (j == y.len() || x[i].0 < y[j].0) {
ret.push((&x[i].0, Some(&x[i].1), None));
i += 1;
} else if j < y.len() && (i == x.len() || x[i].0 > y[j].0) {
2021-03-11 18:30:24 +00:00
ret.push((&y[j].0, None, Some(&y[j].1)));
j += 1;
} else {
unreachable!();
}
}
ret
}