block manager: read_block using old layout versions if necessary

This commit is contained in:
Alex 2023-11-27 11:52:57 +01:00
parent 3ecd14b9f6
commit d6d239fc79
Signed by untrusted user: lx
GPG key ID: 0E496D15096376BE
5 changed files with 50 additions and 9 deletions

View file

@ -264,8 +264,10 @@ impl BlockManager {
F: Fn(DataBlockHeader, ByteStream) -> Fut,
Fut: futures::Future<Output = Result<T, Error>>,
{
let who = self.replication.read_nodes(hash);
let who = self.system.rpc_helper().request_order(&who);
let who = self
.system
.cluster_layout()
.block_read_nodes_of(hash, self.system.rpc_helper());
for node in who.iter() {
let node_id = NodeID::from(*node);

View file

@ -7,6 +7,7 @@ use serde::{Deserialize, Serialize};
use garage_util::data::*;
use super::schema::*;
use crate::rpc_helper::RpcHelper;
#[derive(Debug, Clone, Serialize, Deserialize, Default, PartialEq, Eq)]
pub struct LayoutDigest {
@ -140,6 +141,28 @@ impl LayoutHelper {
.collect()
}
pub fn block_read_nodes_of(&self, position: &Hash, rpc_helper: &RpcHelper) -> Vec<Uuid> {
let mut ret = Vec::with_capacity(12);
let ver_iter = self
.layout()
.versions
.iter()
.rev()
.chain(self.layout().old_versions.iter().rev());
for ver in ver_iter {
if ver.version > self.sync_map_min {
continue;
}
let nodes = ver.nodes_of(position, ver.replication_factor);
for node in rpc_helper.request_order(nodes) {
if !ret.contains(&node) {
ret.push(node);
}
}
}
ret
}
pub(crate) fn write_sets_of(&self, position: &Hash) -> Vec<Vec<Uuid>> {
self.layout()
.versions

View file

@ -18,6 +18,7 @@ impl LayoutHistory {
LayoutHistory {
versions: vec![version],
old_versions: vec![],
update_trackers: Default::default(),
staging: Lww::raw(0, staging),
}
@ -86,11 +87,20 @@ impl LayoutHistory {
.min(&all_nongateway_nodes, min_version);
if self.min_stored() < sync_ack_map_min {
let removed = self.versions.remove(0);
info!("Layout history: pruning old version {}", removed.version);
info!(
"Layout history: moving version {} to old_versions",
removed.version
);
self.old_versions.push(removed);
} else {
break;
}
}
while self.old_versions.len() > OLD_VERSION_COUNT {
let removed = self.old_versions.remove(0);
info!("Layout history: removing old_version {}", removed.version);
}
}
pub(crate) fn clamp_update_trackers(&mut self, nodes: &[Uuid]) {

View file

@ -193,12 +193,18 @@ mod v010 {
use std::collections::BTreeMap;
pub use v09::{LayoutParameters, NodeRole, NodeRoleV, ZoneRedundancy};
pub const OLD_VERSION_COUNT: usize = 5;
/// The history of cluster layouts, with trackers to keep a record
/// of which nodes are up-to-date to current cluster data
#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct LayoutHistory {
/// The versions currently in use in the cluster
pub versions: Vec<LayoutVersion>,
/// At most 5 of the previous versions, not used by the garage_table
/// module, but usefull for the garage_block module to find data blocks
/// that have not yet been moved
pub old_versions: Vec<LayoutVersion>,
/// Update trackers
pub update_trackers: UpdateTrackers,
@ -300,6 +306,7 @@ mod v010 {
};
Self {
versions: vec![version],
old_versions: vec![],
update_trackers: UpdateTrackers {
ack_map: update_tracker.clone(),
sync_map: update_tracker.clone(),

View file

@ -267,7 +267,7 @@ impl RpcHelper {
// When there are errors, we start new requests to compensate.
// Reorder requests to priorize closeness / low latency
let request_order = self.request_order(to);
let request_order = self.request_order(to.iter().copied());
let send_all_at_once = strategy.rs_send_all_at_once.unwrap_or(false);
// Build future for each request
@ -335,7 +335,7 @@ impl RpcHelper {
}
}
pub fn request_order(&self, nodes: &[Uuid]) -> Vec<Uuid> {
pub fn request_order(&self, nodes: impl Iterator<Item = Uuid>) -> Vec<Uuid> {
// Retrieve some status variables that we will use to sort requests
let peer_list = self.0.fullmesh.get_peer_list();
let layout = self.0.layout.read().unwrap();
@ -351,9 +351,8 @@ impl RpcHelper {
// By sorting this vec, we priorize ourself, then nodes in the same zone,
// and within a same zone we priorize nodes with the lowest latency.
let mut nodes = nodes
.iter()
.map(|to| {
let peer_zone = match layout.current().node_role(to) {
let peer_zone = match layout.current().node_role(&to) {
Some(pc) => &pc.zone,
None => "",
};
@ -363,10 +362,10 @@ impl RpcHelper {
.and_then(|pi| pi.avg_ping)
.unwrap_or_else(|| Duration::from_secs(10));
(
*to != self.0.our_node_id,
to != self.0.our_node_id,
peer_zone != our_zone,
peer_avg_ping,
*to,
to,
)
})
.collect::<Vec<_>>();