syncbase/vsync: SyncGroup and DAG changes for SG syncing
* Change SyncGroup storage to support their versioning and being tracked
in DAG and log records.
* Update the SyncGroup join & publish calls to support the asynchronous
transition from known-but-pending SyncGroup to caught-up after sync.
* Add a SyncGroup local state info to track the number of local peer
joiners, and the watchability, remote publishing, and sync pending
states of the SyncGroup.
* Add the retry loop for the SyncGroup publishing to the remote peer.
* Improve error checking of SyncGroup prefixes passed in the Spec.
* Add a 1st-cut HasKey() store util API to bypass the VOM decode. This
will be later further optimized for leveldb to avoid a copy of the
value from the C buffer to the Go buffer.
* Add DAG support to fully prune all nodes for an object including its
current head node.
* Remove temporay solution of the responder adding the initiator to the
SyncGroup joiner list.
* Improve unittest coverage.
Change-Id: I4585e248bd9e15b8b9585ec7b1830f8225686b2a
diff --git a/services/syncbase/server/interfaces/sync.vdl b/services/syncbase/server/interfaces/sync.vdl
index 6dd6f25..80b46a0 100644
--- a/services/syncbase/server/interfaces/sync.vdl
+++ b/services/syncbase/server/interfaces/sync.vdl
@@ -19,15 +19,36 @@
// SyncGroup-related methods.
- // PublishSyncGroup is typically invoked on a "central" peer to publish
- // the SyncGroup.
- PublishSyncGroup(sg SyncGroup) error {access.Write}
+ // PublishSyncGroup is invoked on the SyncGroup name (typically served
+ // by a "central" peer) to publish the SyncGroup. It takes the name of
+ // Syncbase doing the publishing (the publisher) and returns the name
+ // of the Syncbase where the SyncGroup is published (the publishee).
+ // This allows the publisher and the publishee to learn of each other.
+ // When a SyncGroup is published, the publishee is given the SyncGroup
+ // metadata, its current version at the publisher, and the current
+ // SyncGroup generation vector. The generation vector serves as a
+ // checkpoint at the time of publishing. The publishing proceeds
+ // asynchronously, and the publishee learns the SyncGroup history
+ // through the routine p2p sync process and determines when it has
+ // caught up to the level of knowledge at the time of publishing using
+ // the checkpointed generation vector. Until that point, the publishee
+ // locally deems the SyncGroup to be in a pending state and does not
+ // mutate it. Thus it locally rejects SyncGroup joins or updates to
+ // its spec until it is caught up on the SyncGroup history.
+ PublishSyncGroup(publisher string, sg SyncGroup, version string, genvec PrefixGenVector) (string | error) {access.Write}
// JoinSyncGroupAtAdmin is invoked by a prospective SyncGroup member's
// Syncbase on a SyncGroup admin. It checks whether the requestor is
// allowed to join the named SyncGroup, and if so, adds the requestor to
- // the SyncGroup.
- JoinSyncGroupAtAdmin(sgName, joinerName string, myInfo wire.SyncGroupMemberInfo) (SyncGroup | error) {access.Read}
+ // the SyncGroup. It returns a copy of the updated SyncGroup metadata,
+ // its version, and the SyncGroup generation vector at the time of the
+ // join. Similar to the PublishSyncGroup scenario, the joiner at that
+ // point does not have the SyncGroup history and locally deems it to be
+ // in a pending state and does not mutate it. This means it rejects
+ // local updates to the SyncGroup spec or, if it were also an admin on
+ // the SyncGroup, it would reject SyncGroup joins until it is caught up
+ // on the SyncGroup history through p2p sync.
+ JoinSyncGroupAtAdmin(sgName, joinerName string, myInfo wire.SyncGroupMemberInfo) (sg SyncGroup, version string, genvec PrefixGenVector | error) {access.Read}
// BlobSync methods.
@@ -50,3 +71,7 @@
FetchBlobRecipe(br wire.BlobRef) stream<_, ChunkHash> error
FetchChunks() stream<ChunkHash, ChunkData> error
}
+
+error (
+ DupSyncGroupPublish(name string) {"en": "duplicate publish on SyncGroup: {name}"}
+)
