Recover replication baseline from remote when locally pruned #113

Merged
osobh merged 1 commits from fix-replicate-broken-incremental-base into main 2026-08-02 20:59:37 +00:00
2 changed files with 196 additions and 5 deletions
+146 -4
View File
@@ -66,6 +66,23 @@ pub fn replicate_to_cold(
remote_user: &str,
remote_host: &str,
remote_dataset: &str,
) -> Result<()> {
replicate_to_cold_with_state_path(
zfs, dataset, remote_user, remote_host, remote_dataset,
std::path::Path::new(LAST_REPLICATED_PATH),
)
}
/// Same as [`replicate_to_cold`] with the state-file path injectable
/// -- lets tests exercise the baseline-recovery logic against a temp
/// file instead of the real `/var/lib/claw-store/...` path.
pub fn replicate_to_cold_with_state_path(
zfs: &dyn ZfsOps,
dataset: &str,
remote_user: &str,
remote_host: &str,
remote_dataset: &str,
state_path: &std::path::Path,
) -> Result<()> {
let snaps = zfs.list_snapshots(dataset)?;
let latest = match snaps.last().cloned() {
@@ -77,12 +94,64 @@ pub fn replicate_to_cold(
};
// Use the last successfully replicated snapshot as the incremental base.
// Only valid if it still exists in the current snapshot list.
let prev = std::fs::read_to_string(LAST_REPLICATED_PATH)
// Only valid if it still exists in the current snapshot list -- local
// retention (snapshot_retain_hours) can prune it out from under us
// between replication runs (e.g. a daemon restart resets the
// replication tick's 24h timer without resetting the hourly-snapshot
// pruning tick, so a slow/interrupted replication cadence can let the
// recorded snapshot age out locally before it's ever used again).
let recorded = std::fs::read_to_string(state_path)
.ok()
.map(|s| s.trim().to_string())
.filter(|s| !s.is_empty() && snaps.contains(s));
// Recorded snapshot is gone -- don't give up and fall back to a full
// send (which fails outright against a non-empty destination, as
// opposed to just being wasteful). Ask the remote what it actually
// has and find the newest snapshot both sides still share, by tag
// (the `@kind-timestamp` suffix -- dataset paths differ between
// source and destination, e.g. slab/projects vs
// data/archive/tank-projects, but tags are written identically).
let prev = match recorded {
Some(p) => Some(p),
None => {
match zfs.list_remote_snapshots(remote_user, remote_host, remote_dataset) {
Ok(remote_snaps) => {
let remote_tags: std::collections::HashSet<&str> = remote_snaps
.iter()
.filter_map(|s| s.split_once('@').map(|(_, tag)| tag))
.collect();
let fallback = snaps
.iter()
.rev()
.skip(1) // exclude `latest` itself
.find(|s| {
s.split_once('@')
.map(|(_, tag)| remote_tags.contains(tag))
.unwrap_or(false)
})
.cloned();
if let Some(f) = &fallback {
tracing::warn!(
"recorded replication baseline was pruned locally; \
recovered a common snapshot from the remote instead: {}",
f
);
}
fallback
}
Err(e) => {
tracing::warn!(
error = %e,
"could not query remote snapshots to recover a replication baseline; \
falling back to full send"
);
None
}
}
}
};
if prev.as_deref() == Some(latest.as_str()) {
tracing::info!("replication already up to date ({})", latest);
return Ok(());
@@ -99,10 +168,10 @@ pub fn replicate_to_cold(
zfs.send_to_remote(&latest, prev.as_deref(), remote_user, remote_host, remote_dataset)?;
// Record this snapshot as the new baseline for the next incremental send.
if let Some(parent) = std::path::Path::new(LAST_REPLICATED_PATH).parent() {
if let Some(parent) = state_path.parent() {
let _ = std::fs::create_dir_all(parent);
}
std::fs::write(LAST_REPLICATED_PATH, &latest)?;
std::fs::write(state_path, &latest)?;
Ok(())
}
@@ -177,4 +246,77 @@ mod tests {
assert!(!is_sunday_midnight("2026-06-29-0000")); // Monday
assert!(!is_sunday_midnight("2026-06-28-0100")); // Sunday but not midnight
}
/// Unique temp path per test so parallel test runs don't clobber
/// each other's replication-baseline state file.
fn temp_state_path(name: &str) -> std::path::PathBuf {
std::env::temp_dir().join(format!("clawstor-test-last-replicated-{name}-{}", std::process::id()))
}
#[test]
fn replicate_uses_recorded_baseline_when_still_present() {
let zfs = MockZfs::default();
zfs.snapshot("slab/projects", "hourly-1").unwrap();
zfs.snapshot("slab/projects", "hourly-2").unwrap();
let state = temp_state_path("recorded-present");
std::fs::write(&state, "slab/projects@hourly-1").unwrap();
replicate_to_cold_with_state_path(
&zfs, "slab/projects", "user", "host", "remote/ds", &state,
).unwrap();
let sends = zfs.sends();
assert_eq!(sends.len(), 1);
assert_eq!(sends[0].0, "slab/projects@hourly-2");
assert_eq!(sends[0].1.as_deref(), Some("slab/projects@hourly-1"));
let _ = std::fs::remove_file(&state);
}
#[test]
fn replicate_recovers_baseline_from_remote_when_recorded_one_was_pruned() {
let zfs = MockZfs::default();
// Local retention already pruned hourly-1 -- only hourly-2 and
// hourly-3 remain locally. The remote, however, still has
// hourly-2 (it just hasn't received hourly-3 yet).
zfs.snapshot("slab/projects", "hourly-2").unwrap();
zfs.snapshot("slab/projects", "hourly-3").unwrap();
zfs.set_remote_snapshots(vec![
"remote/ds@hourly-1".to_string(),
"remote/ds@hourly-2".to_string(),
]);
let state = temp_state_path("recovers-from-remote");
// Recorded baseline (hourly-1) no longer exists locally.
std::fs::write(&state, "slab/projects@hourly-1").unwrap();
replicate_to_cold_with_state_path(
&zfs, "slab/projects", "user", "host", "remote/ds", &state,
).unwrap();
let sends = zfs.sends();
assert_eq!(sends.len(), 1);
assert_eq!(sends[0].0, "slab/projects@hourly-3");
// Recovered hourly-2 as the base by matching tags against the
// remote's actual snapshot list, NOT a full send.
assert_eq!(sends[0].1.as_deref(), Some("slab/projects@hourly-2"));
let _ = std::fs::remove_file(&state);
}
#[test]
fn replicate_falls_back_to_full_send_when_truly_no_common_snapshot() {
let zfs = MockZfs::default();
zfs.snapshot("slab/projects", "hourly-9").unwrap();
zfs.set_remote_snapshots(vec!["remote/ds@hourly-1".to_string()]);
let state = temp_state_path("no-common-snapshot");
let _ = std::fs::remove_file(&state); // no recorded baseline at all
replicate_to_cold_with_state_path(
&zfs, "slab/projects", "user", "host", "remote/ds", &state,
).unwrap();
let sends = zfs.sends();
assert_eq!(sends.len(), 1);
assert_eq!(sends[0].0, "slab/projects@hourly-9");
assert_eq!(sends[0].1, None, "no common tag exists -- must fall back to full send");
let _ = std::fs::remove_file(&state);
}
}
+50 -1
View File
@@ -13,6 +13,12 @@ pub trait ZfsOps: Send + Sync {
remote_user: &str, remote_host: &str,
remote_dataset: &str) -> Result<()>;
fn clone_snapshot(&self, snapshot: &str, dest_dataset: &str) -> Result<()>;
/// List snapshot names (full `dataset@tag` form, remote-side
/// naming) currently on a remote dataset over SSH. Used to find a
/// real incremental base when the locally-recorded one has been
/// pruned -- see `snapshot::replicate_to_cold`.
fn list_remote_snapshots(&self, remote_user: &str, remote_host: &str,
remote_dataset: &str) -> Result<Vec<String>>;
}
pub struct SystemZfs;
@@ -89,11 +95,48 @@ impl ZfsOps for SystemZfs {
}
Ok(())
}
fn list_remote_snapshots(&self, remote_user: &str, remote_host: &str,
remote_dataset: &str) -> Result<Vec<String>> {
let out = std::process::Command::new("ssh")
.args([
&format!("{remote_user}@{remote_host}"),
"zfs", "list", "-H", "-t", "snapshot", "-o", "name", "-r", remote_dataset,
])
.output()
.context("running ssh zfs list on remote")?;
if !out.status.success() {
bail!("remote zfs list failed: {}", String::from_utf8_lossy(&out.stderr));
}
Ok(String::from_utf8_lossy(&out.stdout)
.lines()
.filter(|l| !l.is_empty())
.map(String::from)
.collect())
}
}
#[derive(Default, Clone)]
pub struct MockZfs {
snapshots: Arc<Mutex<Vec<String>>>,
/// Configurable via `set_remote_snapshots` -- what
/// `list_remote_snapshots` returns, for exercising the
/// baseline-recovery path in `snapshot::replicate_to_cold`.
remote_snapshots: Arc<Mutex<Vec<String>>>,
/// Every `send_to_remote` call, recorded as `(snapshot,
/// incremental_from)`, so tests can assert which base was
/// actually used.
sends: Arc<Mutex<Vec<(String, Option<String>)>>>,
}
impl MockZfs {
pub fn set_remote_snapshots(&self, snaps: Vec<String>) {
*self.remote_snapshots.lock().unwrap() = snaps;
}
pub fn sends(&self) -> Vec<(String, Option<String>)> {
self.sends.lock().unwrap().clone()
}
}
impl ZfsOps for MockZfs {
@@ -115,8 +158,9 @@ impl ZfsOps for MockZfs {
Ok(())
}
fn send_to_remote(&self, _snap: &str, _incr: Option<&str>,
fn send_to_remote(&self, snap: &str, incr: Option<&str>,
_user: &str, _host: &str, _dest: &str) -> Result<()> {
self.sends.lock().unwrap().push((snap.to_string(), incr.map(String::from)));
Ok(())
}
@@ -125,6 +169,11 @@ impl ZfsOps for MockZfs {
.push(format!("{} -> {}", snapshot, dest_dataset));
Ok(())
}
fn list_remote_snapshots(&self, _remote_user: &str, _remote_host: &str,
_remote_dataset: &str) -> Result<Vec<String>> {
Ok(self.remote_snapshots.lock().unwrap().clone())
}
}
#[cfg(test)]