garage/src/table/table.rs

356 lines
9 KiB
Rust
Raw Normal View History

use std::collections::{BTreeMap, HashMap};
use std::sync::Arc;
2022-02-22 12:53:59 +00:00
use std::time::Duration;
2020-04-08 20:00:41 +00:00
2021-10-14 09:50:12 +00:00
use async_trait::async_trait;
use futures::stream::*;
use serde::{Deserialize, Serialize};
use serde_bytes::ByteBuf;
2020-04-24 10:10:01 +00:00
use garage_util::data::*;
use garage_util::error::Error;
2022-02-22 12:53:59 +00:00
use garage_util::metrics::RecordDuration;
2020-04-23 17:05:46 +00:00
2021-10-14 09:50:12 +00:00
use garage_rpc::system::System;
use garage_rpc::*;
2020-04-23 17:05:46 +00:00
2021-05-02 21:13:08 +00:00
use crate::crdt::Crdt;
2021-03-11 15:54:15 +00:00
use crate::data::*;
2021-03-12 20:52:19 +00:00
use crate::gc::*;
2021-03-16 10:43:58 +00:00
use crate::merkle::*;
2021-03-11 15:54:15 +00:00
use crate::replication::*;
use crate::schema::*;
use crate::sync::*;
2020-04-08 20:00:41 +00:00
const TABLE_RPC_TIMEOUT: Duration = Duration::from_secs(10);
2021-10-14 09:50:12 +00:00
pub struct Table<F: TableSchema + 'static, R: TableReplication + 'static> {
2021-03-16 10:43:58 +00:00
pub system: Arc<System>,
pub data: Arc<TableData<F, R>>,
pub merkle_updater: Arc<MerkleUpdater<F, R>>,
2021-03-11 15:54:15 +00:00
pub syncer: Arc<TableSyncer<F, R>>,
2021-10-14 09:50:12 +00:00
endpoint: Arc<Endpoint<TableRpc<F>, Self>>,
2020-04-08 20:00:41 +00:00
}
2020-04-08 21:01:49 +00:00
#[derive(Serialize, Deserialize)]
2021-05-02 21:13:08 +00:00
pub(crate) enum TableRpc<F: TableSchema> {
2020-04-08 21:01:49 +00:00
Ok,
2020-04-08 21:47:34 +00:00
2020-04-09 14:16:27 +00:00
ReadEntry(F::P, F::S),
ReadEntryResponse(Option<ByteBuf>),
2020-04-08 21:47:34 +00:00
// Read range: read all keys in partition P, possibly starting at a certain sort key offset
ReadRange(F::P, Option<F::S>, Option<F::Filter>, usize),
Update(Vec<Arc<ByteBuf>>),
2020-04-08 20:00:41 +00:00
}
impl<F: TableSchema> Rpc for TableRpc<F> {
type Response = Result<TableRpc<F>, Error>;
2021-10-14 09:50:12 +00:00
}
2020-04-18 17:21:34 +00:00
impl<F, R> Table<F, R>
where
F: TableSchema + 'static,
R: TableReplication + 'static,
{
// =============== PUBLIC INTERFACE FUNCTIONS (new, insert, get, etc) ===============
2021-12-14 11:34:01 +00:00
pub fn new(instance: F, replication: R, system: Arc<System>, db: &sled::Db) -> 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/table.rs/Rpc:{}", F::TABLE_NAME));
2020-04-18 17:21:34 +00:00
2021-12-14 11:34:01 +00:00
let data = TableData::new(system.clone(), instance, replication, db);
2021-03-16 10:47:39 +00:00
let merkle_updater = MerkleUpdater::launch(&system.background, data.clone());
2020-04-18 17:21:34 +00:00
2021-10-14 09:50:12 +00:00
let syncer = TableSyncer::launch(system.clone(), data.clone(), merkle_updater.clone());
TableGc::launch(system.clone(), data.clone());
2020-04-18 17:21:34 +00:00
2021-03-12 14:07:23 +00:00
let table = Arc::new(Self {
2021-03-16 10:43:58 +00:00
system,
2021-03-12 14:07:23 +00:00
data,
2021-03-16 10:43:58 +00:00
merkle_updater,
2021-03-12 14:07:23 +00:00
syncer,
2021-10-14 09:50:12 +00:00
endpoint,
2021-03-12 14:07:23 +00:00
});
2021-03-11 15:54:15 +00:00
2021-10-14 09:50:12 +00:00
table.endpoint.set_handler(table.clone());
2020-04-16 12:50:49 +00:00
table
2020-04-08 20:00:41 +00:00
}
2020-04-09 14:16:27 +00:00
pub async fn insert(&self, e: &F::E) -> Result<(), Error> {
let hash = e.partition_key().hash();
2021-03-16 10:43:58 +00:00
let who = self.data.replication.write_nodes(&hash);
2020-04-17 19:59:07 +00:00
//eprintln!("insert who: {:?}", who);
2020-04-08 21:01:49 +00:00
let e_enc = Arc::new(ByteBuf::from(rmp_to_vec_all_named(e)?));
2021-05-02 21:13:08 +00:00
let rpc = TableRpc::<F>::Update(vec![e_enc]);
2021-10-14 09:50:12 +00:00
self.system
.rpc
.try_call_many(
2021-10-14 09:50:12 +00:00
&self.endpoint,
&who[..],
rpc,
2021-10-14 09:50:12 +00:00
RequestStrategy::with_priority(PRIO_NORMAL)
.with_quorum(self.data.replication.write_quorum())
.with_timeout(TABLE_RPC_TIMEOUT),
)
2022-02-22 12:53:59 +00:00
.bound_record_duration(&self.data.metrics.put_request_duration)
.await?;
2022-02-16 13:23:04 +00:00
self.data.metrics.put_request_counter.add(1);
2020-04-08 21:01:49 +00:00
Ok(())
}
2020-04-08 20:00:41 +00:00
pub async fn insert_many(&self, entries: &[F::E]) -> Result<(), Error> {
2022-02-22 12:53:59 +00:00
self.insert_many_internal(entries)
.bound_record_duration(&self.data.metrics.put_request_duration)
.await?;
self.data.metrics.put_request_counter.add(1);
Ok(())
}
2022-02-16 13:23:04 +00:00
2022-02-22 12:53:59 +00:00
async fn insert_many_internal(&self, entries: &[F::E]) -> Result<(), Error> {
2021-04-23 19:42:52 +00:00
let mut call_list: HashMap<_, Vec<_>> = HashMap::new();
for entry in entries.iter() {
let hash = entry.partition_key().hash();
2021-03-16 10:43:58 +00:00
let who = self.data.replication.write_nodes(&hash);
let e_enc = Arc::new(ByteBuf::from(rmp_to_vec_all_named(entry)?));
for node in who {
2021-04-23 19:42:52 +00:00
call_list.entry(node).or_default().push(e_enc.clone());
}
}
let call_futures = call_list.drain().map(|(node, entries)| async move {
2021-05-02 21:13:08 +00:00
let rpc = TableRpc::<F>::Update(entries);
2021-10-14 09:50:12 +00:00
let resp = self
.system
.rpc
.call(
&self.endpoint,
node,
rpc,
RequestStrategy::with_priority(PRIO_NORMAL).with_timeout(TABLE_RPC_TIMEOUT),
)
.await?;
Ok::<_, Error>((node, resp))
});
let mut resps = call_futures.collect::<FuturesUnordered<_>>();
let mut errors = vec![];
while let Some(resp) = resps.next().await {
if let Err(e) = resp {
errors.push(e);
}
}
2021-03-16 10:43:58 +00:00
if errors.len() > self.data.replication.max_write_errors() {
Err(Error::Message("Too many errors".into()))
} else {
Ok(())
}
}
2020-04-16 12:50:49 +00:00
pub async fn get(
self: &Arc<Self>,
partition_key: &F::P,
sort_key: &F::S,
) -> Result<Option<F::E>, Error> {
2022-02-22 12:53:59 +00:00
let res = self.get_internal(partition_key, sort_key)
.bound_record_duration(&self.data.metrics.get_request_duration)
.await?;
self.data.metrics.get_request_counter.add(1);
Ok(res)
}
2022-02-16 13:23:04 +00:00
2022-02-22 12:53:59 +00:00
async fn get_internal(
self: &Arc<Self>,
partition_key: &F::P,
sort_key: &F::S,
) -> Result<Option<F::E>, Error> {
2020-04-09 14:16:27 +00:00
let hash = partition_key.hash();
2021-03-16 10:43:58 +00:00
let who = self.data.replication.read_nodes(&hash);
2020-04-08 20:00:41 +00:00
2021-05-02 21:13:08 +00:00
let rpc = TableRpc::<F>::ReadEntry(partition_key.clone(), sort_key.clone());
let resps = self
2021-10-14 09:50:12 +00:00
.system
.rpc
.try_call_many(
2021-10-14 09:50:12 +00:00
&self.endpoint,
&who[..],
rpc,
2021-10-14 09:50:12 +00:00
RequestStrategy::with_priority(PRIO_NORMAL)
.with_quorum(self.data.replication.read_quorum())
.with_timeout(TABLE_RPC_TIMEOUT)
.interrupt_after_quorum(true),
)
2020-04-08 21:01:49 +00:00
.await?;
2020-04-08 21:47:34 +00:00
let mut ret = None;
2020-04-09 18:58:39 +00:00
let mut not_all_same = false;
2020-04-08 21:01:49 +00:00
for resp in resps {
2021-05-02 21:13:08 +00:00
if let TableRpc::ReadEntryResponse(value) = resp {
if let Some(v_bytes) = value {
2021-03-11 15:54:15 +00:00
let v = self.data.decode_entry(v_bytes.as_slice())?;
2020-04-08 21:47:34 +00:00
ret = match ret {
None => Some(v),
Some(mut x) => {
2020-04-09 21:45:07 +00:00
if x != v {
2020-04-09 18:58:39 +00:00
not_all_same = true;
2020-04-09 21:45:07 +00:00
x.merge(&v);
2020-04-09 18:58:39 +00:00
}
2020-04-08 21:47:34 +00:00
Some(x)
}
}
2020-04-08 21:01:49 +00:00
}
2020-04-08 21:47:34 +00:00
} else {
2021-04-23 19:42:52 +00:00
return Err(Error::Message("Invalid return value to read".to_string()));
2020-04-08 21:01:49 +00:00
}
}
2020-04-09 18:58:39 +00:00
if let Some(ret_entry) = &ret {
if not_all_same {
let self2 = self.clone();
let ent2 = ret_entry.clone();
2021-03-16 10:43:58 +00:00
self.system
2020-04-16 12:50:49 +00:00
.background
.spawn_cancellable(async move { self2.repair_on_read(&who[..], ent2).await });
2020-04-09 18:58:39 +00:00
}
}
2022-02-16 13:23:04 +00:00
2020-04-08 21:47:34 +00:00
Ok(ret)
2020-04-08 21:01:49 +00:00
}
pub async fn get_range(
self: &Arc<Self>,
partition_key: &F::P,
begin_sort_key: Option<F::S>,
filter: Option<F::Filter>,
limit: usize,
) -> Result<Vec<F::E>, Error> {
2022-02-22 12:53:59 +00:00
let res = self.get_range_internal(partition_key, begin_sort_key, filter, limit)
.bound_record_duration(&self.data.metrics.get_request_duration)
.await?;
self.data.metrics.get_request_counter.add(1);
Ok(res)
}
2022-02-16 13:23:04 +00:00
2022-02-22 12:53:59 +00:00
async fn get_range_internal(
self: &Arc<Self>,
partition_key: &F::P,
begin_sort_key: Option<F::S>,
filter: Option<F::Filter>,
limit: usize,
) -> Result<Vec<F::E>, Error> {
let hash = partition_key.hash();
2021-03-16 10:43:58 +00:00
let who = self.data.replication.read_nodes(&hash);
2021-05-02 21:13:08 +00:00
let rpc = TableRpc::<F>::ReadRange(partition_key.clone(), begin_sort_key, filter, limit);
let resps = self
2021-10-14 09:50:12 +00:00
.system
.rpc
.try_call_many(
2021-10-14 09:50:12 +00:00
&self.endpoint,
&who[..],
rpc,
2021-10-14 09:50:12 +00:00
RequestStrategy::with_priority(PRIO_NORMAL)
.with_quorum(self.data.replication.read_quorum())
.with_timeout(TABLE_RPC_TIMEOUT)
.interrupt_after_quorum(true),
)
.await?;
let mut ret = BTreeMap::new();
let mut to_repair = BTreeMap::new();
for resp in resps {
2021-05-02 21:13:08 +00:00
if let TableRpc::Update(entries) = resp {
for entry_bytes in entries.iter() {
2021-03-11 15:54:15 +00:00
let entry = self.data.decode_entry(entry_bytes.as_slice())?;
let entry_key = self.data.tree_key(entry.partition_key(), entry.sort_key());
match ret.remove(&entry_key) {
None => {
ret.insert(entry_key, Some(entry));
}
Some(Some(mut prev)) => {
let must_repair = prev != entry;
prev.merge(&entry);
if must_repair {
to_repair.insert(entry_key.clone(), Some(prev.clone()));
}
ret.insert(entry_key, Some(prev));
}
Some(None) => unreachable!(),
}
}
}
}
if !to_repair.is_empty() {
let self2 = self.clone();
2021-03-16 10:43:58 +00:00
self.system.background.spawn_cancellable(async move {
for (_, v) in to_repair.iter_mut() {
self2.repair_on_read(&who[..], v.take().unwrap()).await?;
}
Ok(())
});
}
let ret_vec = ret
.iter_mut()
.take(limit)
.map(|(_k, v)| v.take().unwrap())
.collect::<Vec<_>>();
Ok(ret_vec)
}
// =============== UTILITY FUNCTION FOR CLIENT OPERATIONS ===============
async fn repair_on_read(&self, who: &[Uuid], what: F::E) -> Result<(), Error> {
2020-04-16 12:50:49 +00:00
let what_enc = Arc::new(ByteBuf::from(rmp_to_vec_all_named(&what)?));
2021-10-14 09:50:12 +00:00
self.system
.rpc
2020-04-18 17:21:34 +00:00
.try_call_many(
2021-10-14 09:50:12 +00:00
&self.endpoint,
2021-04-23 19:42:52 +00:00
who,
2021-05-02 21:13:08 +00:00
TableRpc::<F>::Update(vec![what_enc]),
2021-10-14 09:50:12 +00:00
RequestStrategy::with_priority(PRIO_NORMAL)
.with_quorum(who.len())
.with_timeout(TABLE_RPC_TIMEOUT),
2020-04-18 17:21:34 +00:00
)
2020-04-16 12:50:49 +00:00
.await?;
Ok(())
}
}
#[async_trait]
impl<F, R> EndpointHandler<TableRpc<F>> for Table<F, R>
where
F: TableSchema + 'static,
R: TableReplication + 'static,
{
async fn handle(
self: &Arc<Self>,
msg: &TableRpc<F>,
_from: NodeID,
) -> Result<TableRpc<F>, Error> {
2020-04-08 21:01:49 +00:00
match msg {
2021-05-02 21:13:08 +00:00
TableRpc::ReadEntry(key, sort_key) => {
2021-03-11 15:54:15 +00:00
let value = self.data.read_entry(key, sort_key)?;
2021-05-02 21:13:08 +00:00
Ok(TableRpc::ReadEntryResponse(value))
2020-04-08 21:01:49 +00:00
}
2021-05-02 21:13:08 +00:00
TableRpc::ReadRange(key, begin_sort_key, filter, limit) => {
2021-03-11 15:54:15 +00:00
let values = self.data.read_range(key, begin_sort_key, filter, *limit)?;
2021-05-02 21:13:08 +00:00
Ok(TableRpc::Update(values))
}
2021-05-02 21:13:08 +00:00
TableRpc::Update(pairs) => {
2021-03-11 15:54:15 +00:00
self.data.update_many(pairs)?;
2021-05-02 21:13:08 +00:00
Ok(TableRpc::Ok)
2020-04-08 21:01:49 +00:00
}
m => Err(Error::unexpected_rpc_message(m)),
2020-04-08 21:01:49 +00:00
}
}
2020-04-08 20:00:41 +00:00
}