tka/sync: send checkpoints to ensure far-behind nodes can catch up

Previously there was a mismatch between how nodes store AUMs and what
the control plane would offer during sync:

- Client compaction: Nodes aggressively compact their TKA state -- they
  keep the last 24 AUMs, every AUM received in the last two weeks, and
  then everything from there back to the last checkpoint. Depending on
  when it compacts, a node may only have ~50 AUMs.
- Exponential sampling: To save bandwidth, the control plane would send
  a SyncOffer containing ancestors at exponentially increasing intervals
  (4th, 16th, 64th, 256th...).

If a node has been offline for too long, the exponential sampling skips
the node's smaller window. When the SyncOffer and local state are disjoint,
the node cannot find a common ancestor to use for synchronisation.
It enters a failure loop where it keeps polling for new TKA state, but
it cannot catch up and has an increasingly-outdated view of the tailnet.

This patch replaces the exponential sampling with a SyncOffer that sends
every checkpoint ancestor of the current HEAD. Since every node is
guaranteed to keep at least one checkpoint after compaction, we're more
likely to have an intersection for the sync process.

This patch also increases `maxSyncHeadIntersectionIter`, which in
practice means the control plane will send every checkpoint in the
current chain. This means all affected nodes will be able to find an
intersection and catch up immediately, without requiring a client update.

It's still possible for a node to be unable to sync, but these edge cases
become less likely with this change. (For example, if a node is 1000+ AUMs
behind, or if it creates a local branch and then compacts away the
intersection with the main chain.)

This patch includes a regression test with synthetic data, and I
verified the fix with customer data.

Updates https://github.com/tailscale/corp/issues/40404

Change-Id: I2174011bb23a2b5972f6d1591aadcc016e3cba35
Signed-off-by: Alex Chan <alexc@tailscale.com>

tka/sync.go

@@ -60,18 +60,6 @@ func FromSyncOffer(offer SyncOffer) (head string, ancestors []string, err error)
 	return string(headBytes), ancestors, nil
 }
 
-const (
-	// The starting number of AUMs to skip when listing
-	// ancestors in a SyncOffer.
-	ancestorsSkipStart = 4
-
-	// How many bits to advance the skip count when listing
-	// ancestors in a SyncOffer.
-	//
-	// 2 bits, so (4<<2), so after skipping 4 it skips 16.
-	ancestorsSkipShift = 2
-)
-
 // SyncOffer returns an abbreviated description of the current AUM
 // chain, which can be used to synchronize with another (untrusted)
 // Authority instance.
@@ -92,20 +80,10 @@ func (a *Authority) SyncOffer(storage Chonk) (SyncOffer, error) {
 		Ancestors: make([]AUMHash, 0, 6), // 6 chosen arbitrarily.
 	}
 
-	// We send some subset of our ancestors to help the remote
-	// find a more-recent 'head intersection'.
-	// The number of AUMs between each ancestor entry gets
-	// exponentially larger.
-	var (
-		skipAmount uint64  = ancestorsSkipStart
-		curs       AUMHash = a.Head()
-	)
-	for i := range uint64(maxSyncHeadIntersectionIter) {
-		if i > 0 && (i%skipAmount) == 0 {
-			out.Ancestors = append(out.Ancestors, curs)
-			skipAmount = skipAmount << ancestorsSkipShift
-		}
-
+	// We send all our checkpoints to help the remote find a
+	// more-recent 'head intersection'.
+	curs := a.Head()
+	for range maxSyncHeadIntersectionIter {
 		parent, err := storage.AUM(curs)
 		if err != nil {
 			if err != os.ErrNotExist {
@@ -118,6 +96,11 @@ func (a *Authority) SyncOffer(storage Chonk) (SyncOffer, error) {
 		if parent.Hash() == oldest {
 			break
 		}
+
+		if parent.MessageKind == AUMCheckpoint {
+			out.Ancestors = append(out.Ancestors, curs)
+		}
+
 		copy(curs[:], parent.PrevAUMHash)
 	}