feat(pageserver): support detaching from multiple ancestors

Signed-off-by: Alex Chi Z <chi@neon.tech>
This commit is contained in:
Alex Chi Z
2025-03-10 14:54:42 -04:00
parent fb1957936c
commit 7d6c9d0557
5 changed files with 219 additions and 21 deletions

View File

@@ -300,9 +300,8 @@ impl TimelineMetadata {
/// Returns true if anything was changed
pub fn detach_from_ancestor(&mut self, branchpoint: &(TimelineId, Lsn)) {
if let Some(ancestor) = self.body.ancestor_timeline {
assert_eq!(ancestor, branchpoint.0);
}
// Detaching from ancestor now doesn't always detach directly to the direct ancestor, but we
// ensure the LSN is the same. So we don't check the timeline ID.
if self.body.ancestor_lsn != Lsn(0) {
assert_eq!(self.body.ancestor_lsn, branchpoint.1);
}

View File

@@ -1957,7 +1957,13 @@ impl TenantManager {
.map_err(Error::NotFound)?;
let resp = timeline
.detach_from_ancestor_and_reparent(&tenant, prepared, ctx)
.detach_from_ancestor_and_reparent(
&tenant,
prepared,
attempt.ancestor_timeline_id,
attempt.ancestor_lsn,
ctx,
)
.await?;
let mut slot_guard = slot_guard;

View File

@@ -5406,9 +5406,19 @@ impl Timeline {
self: &Arc<Timeline>,
tenant: &crate::tenant::Tenant,
prepared: detach_ancestor::PreparedTimelineDetach,
ancestor_timeline_id: TimelineId,
ancestor_lsn: Lsn,
ctx: &RequestContext,
) -> Result<detach_ancestor::DetachingAndReparenting, detach_ancestor::Error> {
detach_ancestor::detach_and_reparent(self, tenant, prepared, ctx).await
detach_ancestor::detach_and_reparent(
self,
tenant,
prepared,
ancestor_timeline_id,
ancestor_lsn,
ctx,
)
.await
}
/// Final step which unblocks the GC.

View File

@@ -29,7 +29,7 @@ pub(crate) enum Error {
#[error("no ancestors")]
NoAncestor,
#[error("too many ancestors")]
#[error("the branch has more than 1 ancestor and cannot be fast-path detached")]
TooManyAncestors,
#[error("shutting down, please retry later")]
@@ -147,7 +147,8 @@ impl Default for Options {
#[derive(Debug)]
pub(crate) struct Attempt {
pub(crate) timeline_id: TimelineId,
pub(crate) ancestor_timeline_id: TimelineId,
pub(crate) ancestor_lsn: Lsn,
_guard: completion::Completion,
gate_entered: Option<utils::sync::gate::GateGuard>,
}
@@ -172,20 +173,24 @@ pub(super) async fn prepare(
) -> Result<Progress, Error> {
use Error::*;
let Some((ancestor, ancestor_lsn)) = detached
let Some((mut ancestor, mut ancestor_lsn)) = detached
.ancestor_timeline
.as_ref()
.map(|tl| (tl.clone(), detached.ancestor_lsn))
else {
let ancestor_id;
let ancestor_lsn;
let still_in_progress = {
let accessor = detached.remote_client.initialized_upload_queue()?;
// we are safe to inspect the latest uploaded, because we can only witness this after
// restart is complete and ancestor is no more.
let latest = accessor.latest_uploaded_index_part();
if latest.lineage.detached_previous_ancestor().is_none() {
let Some((id, lsn)) = latest.lineage.detached_previous_ancestor() else {
return Err(NoAncestor);
};
ancestor_id = id;
ancestor_lsn = lsn;
latest
.gc_blocking
@@ -196,7 +201,8 @@ pub(super) async fn prepare(
if still_in_progress {
// gc is still blocked, we can still reparent and complete.
// we are safe to reparent remaining, because they were locked in in the beginning.
let attempt = continue_with_blocked_gc(detached, tenant).await?;
let attempt =
continue_with_blocked_gc(detached, tenant, ancestor_id, ancestor_lsn).await?;
// because the ancestor of detached is already set to none, we have published all
// of the layers, so we are still "prepared."
@@ -224,13 +230,26 @@ pub(super) async fn prepare(
check_no_archived_children_of_ancestor(tenant, detached, &ancestor, ancestor_lsn)?;
if ancestor.ancestor_timeline.is_some() {
// non-technical requirement; we could flatten N ancestors just as easily but we chose
// not to, at least initially
return Err(TooManyAncestors);
// If the ancestor has an ancestor, we might be able to fast-path detach it if the current ancestor does not have any data written/used by the detaching timeline.
while let Some(ancestor_of_ancestor) = ancestor.ancestor_timeline.clone() {
if ancestor_lsn != ancestor.ancestor_lsn {
// non-technical requirement; we could flatten N ancestors just as easily but we chose
// not to, at least initially
return Err(TooManyAncestors);
}
// Use the ancestor of the ancestor as the new ancestor (only when the ancestor LSNs are the same)
ancestor_lsn = ancestor.ancestor_lsn; // Get the LSN first before resetting the `ancestor` variable
ancestor = ancestor_of_ancestor;
check_no_archived_children_of_ancestor(tenant, detached, &ancestor, ancestor_lsn)?;
}
let attempt = start_new_attempt(detached, tenant).await?;
tracing::info!(
"attempt to detach the timeline from the ancestor: {}@{}",
ancestor.timeline_id,
ancestor_lsn
);
let attempt = start_new_attempt(detached, tenant, ancestor.timeline_id, ancestor_lsn).await?;
utils::pausable_failpoint!("timeline-detach-ancestor::before_starting_after_locking-pausable");
@@ -450,8 +469,13 @@ pub(super) async fn prepare(
Ok(Progress::Prepared(attempt, prepared))
}
async fn start_new_attempt(detached: &Timeline, tenant: &Tenant) -> Result<Attempt, Error> {
let attempt = obtain_exclusive_attempt(detached, tenant)?;
async fn start_new_attempt(
detached: &Timeline,
tenant: &Tenant,
ancestor_timeline_id: TimelineId,
ancestor_lsn: Lsn,
) -> Result<Attempt, Error> {
let attempt = obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)?;
// insert the block in the index_part.json, if not already there.
let _dont_care = tenant
@@ -466,13 +490,23 @@ async fn start_new_attempt(detached: &Timeline, tenant: &Tenant) -> Result<Attem
Ok(attempt)
}
async fn continue_with_blocked_gc(detached: &Timeline, tenant: &Tenant) -> Result<Attempt, Error> {
async fn continue_with_blocked_gc(
detached: &Timeline,
tenant: &Tenant,
ancestor_timeline_id: TimelineId,
ancestor_lsn: Lsn,
) -> Result<Attempt, Error> {
// FIXME: it would be nice to confirm that there is an in-memory version, since we've just
// verified there is a persistent one?
obtain_exclusive_attempt(detached, tenant)
obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)
}
fn obtain_exclusive_attempt(detached: &Timeline, tenant: &Tenant) -> Result<Attempt, Error> {
fn obtain_exclusive_attempt(
detached: &Timeline,
tenant: &Tenant,
ancestor_timeline_id: TimelineId,
ancestor_lsn: Lsn,
) -> Result<Attempt, Error> {
use Error::{OtherTimelineDetachOngoing, ShuttingDown};
// ensure we are the only active attempt for this tenant
@@ -493,6 +527,8 @@ fn obtain_exclusive_attempt(detached: &Timeline, tenant: &Tenant) -> Result<Atte
Ok(Attempt {
timeline_id: detached.timeline_id,
ancestor_timeline_id,
ancestor_lsn,
_guard: guard,
gate_entered: Some(_gate_entered),
})
@@ -795,6 +831,8 @@ pub(super) async fn detach_and_reparent(
detached: &Arc<Timeline>,
tenant: &Tenant,
prepared: PreparedTimelineDetach,
ancestor_timeline_id: TimelineId,
ancestor_lsn: Lsn,
_ctx: &RequestContext,
) -> Result<DetachingAndReparenting, Error> {
let PreparedTimelineDetach { layers } = prepared;
@@ -822,7 +860,30 @@ pub(super) async fn detach_and_reparent(
"cannot (detach? reparent)? complete if the operation is not still ongoing"
);
let ancestor = match (detached.ancestor_timeline.as_ref(), recorded_branchpoint) {
let ancestor_to_detach = match detached.ancestor_timeline.as_ref() {
Some(mut ancestor) => {
while ancestor.timeline_id != ancestor_timeline_id {
match ancestor.ancestor_timeline.as_ref() {
Some(found) => {
if ancestor_lsn != ancestor.ancestor_lsn {
return Err(Error::DetachReparent(anyhow::anyhow!(
"cannot find the ancestor timeline to detach from: wrong ancestor lsn"
)));
}
ancestor = found;
}
None => {
return Err(Error::DetachReparent(anyhow::anyhow!(
"cannot find the ancestor timeline to detach from"
)));
}
}
}
Some(ancestor)
}
None => None,
};
let ancestor = match (ancestor_to_detach, recorded_branchpoint) {
(Some(ancestor), None) => {
assert!(
!layers.is_empty(),

View File

@@ -342,6 +342,128 @@ def test_ancestor_detach_reparents_earlier(neon_env_builder: NeonEnvBuilder):
wait_timeline_detail_404(client, env.initial_tenant, env.initial_timeline)
def test_ancestor_detach_two_level_ancestors(neon_env_builder: NeonEnvBuilder):
"""
old main -------|---------X--------->
| | |
| | +-> after
| +--X empty snapshot branch
| |
| +-> branch-to-detach
|
+-> reparented
Ends up as:
old main -------|---------X--------->
| |
| +-> after
+--> empty snapshot branch
new main -------|--------------> branch-to-detach
|
+-> reparented
"""
env = neon_env_builder.init_start()
env.pageserver.allowed_errors.extend(SHUTDOWN_ALLOWED_ERRORS)
client = env.pageserver.http_client()
with env.endpoints.create_start("main", tenant_id=env.initial_tenant) as ep:
ep.safe_psql("CREATE TABLE foo (i BIGINT);")
ep.safe_psql("CREATE TABLE audit AS SELECT 1 as starts;")
branchpoint_pipe = wait_for_last_flush_lsn(
env, ep, env.initial_tenant, env.initial_timeline
)
ep.safe_psql("INSERT INTO foo SELECT i::bigint FROM generate_series(0, 8191) g(i);")
branchpoint_x = wait_for_last_flush_lsn(env, ep, env.initial_tenant, env.initial_timeline)
client.timeline_checkpoint(env.initial_tenant, env.initial_timeline)
ep.safe_psql("INSERT INTO foo SELECT i::bigint FROM generate_series(8192, 16383) g(i);")
wait_for_last_flush_lsn(env, ep, env.initial_tenant, env.initial_timeline)
reparented = env.create_branch(
"reparented", ancestor_branch_name="main", ancestor_start_lsn=branchpoint_pipe
)
snapshot_branchpoint = env.create_branch(
"snapshot_branchpoint", ancestor_branch_name="main", ancestor_start_lsn=branchpoint_x
)
branch_to_detach = env.create_branch(
"branch_to_detach", ancestor_branch_name="snapshot_branchpoint", ancestor_start_lsn=branchpoint_x
)
after = env.create_branch("after", ancestor_branch_name="main", ancestor_start_lsn=None)
all_reparented = client.detach_ancestor(env.initial_tenant, branch_to_detach)
assert set(all_reparented) == {reparented, snapshot_branchpoint}
env.pageserver.quiesce_tenants()
# checking the ancestor after is much faster than waiting for the endpoint not start
expected_result = [
("main", env.initial_timeline, None, 16384, 1),
("after", after, env.initial_timeline, 16384, 1),
("snapshot_branchpoint", snapshot_branchpoint, branch_to_detach, 16384, 1), # not correct
("branch_to_detach", branch_to_detach, None, 8192, 1),
("reparented", reparented, env.initial_timeline, 0, 1),
]
assert isinstance(env.pageserver_remote_storage, LocalFsStorage)
for branch_name, queried_timeline, expected_ancestor, _, _ in expected_result:
details = client.timeline_detail(env.initial_tenant, queried_timeline)
ancestor_timeline_id = details["ancestor_timeline_id"]
if expected_ancestor is None:
assert ancestor_timeline_id is None
else:
assert TimelineId(ancestor_timeline_id) == expected_ancestor, f"when checking branch {branch_name}, mapping={expected_result}"
index_part = env.pageserver_remote_storage.index_content(
env.initial_tenant, queried_timeline
)
lineage = index_part["lineage"]
assert lineage is not None
assert lineage.get("reparenting_history_overflown", "false") == "false"
if queried_timeline == branch_to_detach:
original_ancestor = lineage["original_ancestor"]
assert original_ancestor is not None
assert original_ancestor[0] == str(env.initial_timeline)
assert original_ancestor[1] == str(branchpoint_x)
# this does not contain Z in the end, so fromisoformat accepts it
# it is to be in line with the deletion timestamp.. well, almost.
when = original_ancestor[2][:26]
when_ts = datetime.datetime.fromisoformat(when)
assert when_ts < datetime.datetime.now()
assert len(lineage.get("reparenting_history", [])) == 0
elif expected_ancestor == branch_to_detach:
assert len(lineage.get("original_ancestor", [])) == 0
assert lineage["reparenting_history"] == [str(env.initial_timeline)]
else:
assert len(lineage.get("original_ancestor", [])) == 0
assert len(lineage.get("reparenting_history", [])) == 0
for name, _, _, rows, starts in expected_result:
with env.endpoints.create_start(name, tenant_id=env.initial_tenant) as ep:
assert ep.safe_psql("SELECT count(*) FROM foo;")[0][0] == rows
assert ep.safe_psql(f"SELECT count(*) FROM audit WHERE starts = {starts}")[0][0] == 1
# delete the timelines to confirm detach actually worked
client.timeline_delete(env.initial_tenant, after)
wait_timeline_detail_404(client, env.initial_tenant, after)
client.timeline_delete(env.initial_tenant, env.initial_timeline)
wait_timeline_detail_404(client, env.initial_tenant, env.initial_timeline)
def test_detached_receives_flushes_while_being_detached(neon_env_builder: NeonEnvBuilder):
"""
Makes sure that the timeline is able to receive writes through-out the detach process.