snapshots

State Sync Snapshotting

The snapshots package implements automatic support for Tendermint state sync in Cosmos SDK-based applications. State sync allows a new node joining a network to simply fetch a recent snapshot of the application state instead of fetching and applying all historical blocks. This can reduce the time needed to join the network by several orders of magnitude (e.g. weeks to minutes), but the node will not contain historical data from previous heights.

This document describes the Cosmos SDK implementation of the ABCI state sync interface, for more information on Tendermint state sync in general see:

• Tendermint Core State Sync for Developers
• ABCI State Sync Spec
• ABCI State Sync Method/Type Reference

Overview

For an overview of how Cosmos SDK state sync is set up and configured by developers and end-users, see the Cosmos SDK State Sync Guide.

Briefly, the Cosmos SDK takes state snapshots at regular height intervals given by state-sync.snapshot-interval and stores them as binary files in the filesystem under <node_home>/data/snapshots/, with metadata in a LevelDB database <node_home>/data/snapshots/metadata.db. The number of recent snapshots to keep are given by state-sync.snapshot-keep-recent.

Snapshots are taken asynchronously, i.e. new blocks will be applied concurrently with snapshots being taken. This is possible because IAVL supports querying immutable historical heights. However, this requires state-sync.snapshot-interval to be a multiple of pruning-keep-every, to prevent a height from being removed while it is being snapshotted.

When a remote node is state syncing, Tendermint calls the ABCI method ListSnapshots to list available local snapshots and LoadSnapshotChunk to load a binary snapshot chunk. When the local node is being state synced, Tendermint calls OfferSnapshot to offer a discovered remote snapshot to the local application and ApplySnapshotChunk to apply a binary snapshot chunk to the local application. See the resources linked above for more details on these methods and how Tendermint performs state sync.

The Cosmos SDK does not currently do any incremental verification of snapshots during restoration, i.e. only after the entire snapshot has been restored will Tendermint compare the app hash against the trusted hash from the chain. Cosmos SDK snapshots and chunks do contain hashes as checksums to guard against IO corruption and non-determinism, but these are not tied to the chain state and can be trivially forged by an adversary. This was considered out of scope for the initial implementation, but can be added later without changes to the ABCI state sync protocol.

Snapshot Metadata

The ABCI Protobuf type for a snapshot is listed below (refer to the ABCI spec for field details):

message Snapshot {
  uint64 height   = 1;  // The height at which the snapshot was taken
  uint32 format   = 2;  // The application-specific snapshot format
  uint32 chunks   = 3;  // Number of chunks in the snapshot
  bytes  hash     = 4;  // Arbitrary snapshot hash, equal only if identical
  bytes  metadata = 5;  // Arbitrary application metadata
}

Because the metadata field is application-specific, the Cosmos SDK uses a similar type cosmos.base.snapshots.v1beta1.Snapshot with its own metadata representation:

// Snapshot contains Tendermint state sync snapshot info.
message Snapshot {
  uint64   height   = 1;
  uint32   format   = 2;
  uint32   chunks   = 3;
  bytes    hash     = 4;
  Metadata metadata = 5 [(gogoproto.nullable) = false];
}

// Metadata contains SDK-specific snapshot metadata.
message Metadata {
  repeated bytes chunk_hashes = 1; // SHA-256 chunk hashes
}

The format is currently 1, defined in snapshots.types.CurrentFormat. This must be increased whenever the binary snapshot format changes, and it may be useful to support past formats in newer versions.

The hash is a SHA-256 hash of the entire binary snapshot, used to guard against IO corruption and non-determinism across nodes. Note that this is not tied to the chain state, and can be trivially forged (but Tendermint will always compare the final app hash against the chain app hash). Similarly, the chunk_hashes are SHA-256 checksums of each binary chunk.

The metadata field is Protobuf-serialized before it is placed into the ABCI snapshot.

Snapshot Format

The current version 1 snapshot format is a zlib-compressed, length-prefixed Protobuf stream of cosmos.base.store.v1beta1.SnapshotItem messages, split into chunks at exact 10 MB byte boundaries.

// SnapshotItem is an item contained in a rootmulti.Store snapshot.
message SnapshotItem {
  // item is the specific type of snapshot item.
  oneof item {
    SnapshotStoreItem store = 1;
    SnapshotIAVLItem  iavl  = 2 [(gogoproto.customname) = "IAVL"];
  }
}

// SnapshotStoreItem contains metadata about a snapshotted store.
message SnapshotStoreItem {
  string name = 1;
}

// SnapshotIAVLItem is an exported IAVL node.
message SnapshotIAVLItem {
  bytes key     = 1;
  bytes value   = 2;
  int64 version = 3;
  int32 height  = 4;
}

Snapshots are generated by rootmulti.Store.Snapshot() as follows:

1. Set up a protoio.NewDelimitedWriter that writes length-prefixed serialized SnapshotItem Protobuf messages.
   1. Iterate over each IAVL store in lexicographical order by store name.
   2. Emit a SnapshotStoreItem containing the store name.
   3. Start an IAVL export for the store using iavl.ImmutableTree.Export().
   4. Iterate over each IAVL node.
   5. Emit a SnapshotIAVLItem for the IAVL node.
2. Pass the serialized Protobuf output stream to a zlib compression writer.
3. Split the zlib output stream into chunks at exactly every 10th megabyte.

Snapshots are restored via rootmulti.Store.Restore() as the inverse of the above, using iavl.MutableTree.Import() to reconstruct each IAVL tree.

Snapshot Storage

Snapshot storage is managed by snapshots.Store, with metadata in a db.DB database and binary chunks in the filesystem. Note that this is only used to store locally taken snapshots that are being offered to other nodes. When the local node is being state synced, Tendermint will take care of buffering and storing incoming snapshot chunks before they are applied to the application.

Metadata is generally stored in a LevelDB database at <node_home>/data/snapshots/metadata.db. It contains serialized cosmos.base.snapshots.v1beta1.Snapshot Protobuf messages with a key given by the concatenation of a key prefix, the big-endian height, and the big-endian format. Chunk data is stored as regular files under <node_home>/data/snapshots/<height>/<format>/<chunk>.

The snapshots.Store API is based on streaming IO, and integrates easily with the snapshots.types.Snapshotter snapshot/restore interface implemented by rootmulti.Store. The Store.Save() method stores a snapshot given as a <- chan io.ReadCloser channel of binary chunk streams, and Store.Load() loads the snapshot as a channel of binary chunk streams -- the same stream types used by Snapshotter.Snapshot() and Snapshotter.Restore() to take and restore snapshots using streaming IO.

The store also provides many other methods such as List() to list stored snapshots, LoadChunk() to load a single snapshot chunk, and Prune() to prune old snapshots.

Taking Snapshots

snapshots.Manager is a high-level snapshot manager that integrates a snapshots.types.Snapshotter (i.e. the rootmulti.Store snapshot functionality) and a snapshots.Store, providing an API that maps easily onto the ABCI state sync API. The Manager will also make sure only one operation is in progress at a time, e.g. to prevent multiple snapshots being taken concurrently.

During BaseApp.Commit, once a state transition has been committed, the height is checked against the state-sync.snapshot-interval setting. If the committed height should be snapshotted, a goroutine BaseApp.snapshot() is spawned that calls snapshots.Manager.Create() to create the snapshot.

Manager.Create() will do some basic pre-flight checks, and then start generating a snapshot by calling rootmulti.Store.Snapshot(). The chunk stream is passed into snapshots.Store.Save(), which stores the chunks in the filesystem and records the snapshot metadata in the snapshot database.

Once the snapshot has been generated, BaseApp.snapshot() then removes any old snapshots based on the state-sync.snapshot-keep-recent setting.

Serving Snapshots

When a remote node is discovering snapshots for state sync, Tendermint will call the ListSnapshots ABCI method to list the snapshots present on the local node. This is dispatched to snapshots.Manager.List(), which in turn dispatches to snapshots.Store.List().

When a remote node is fetching snapshot chunks during state sync, Tendermint will call the LoadSnapshotChunk ABCI method to fetch a chunk from the local node. This dispatches to snapshots.Manager.LoadChunk(), which in turn dispatches to snapshots.Store.LoadChunk().

Restoring Snapshots

When the operator has configured the local Tendermint node to run state sync (see the resources listed in the introduction for details on Tendermint state sync), it will discover snapshots across the P2P network and offer their metadata in turn to the local application via the OfferSnapshot ABCI call.

BaseApp.OfferSnapshot() attempts to start a restore operation by calling snapshots.Manager.Restore(). This may fail, e.g. if the snapshot format is unknown (it may have been generated by a different version of the Cosmos SDK), in which case Tendermint will offer other discovered snapshots.

If the snapshot is accepted, Manager.Restore() will record that a restore operation is in progress, and spawn a separate goroutine that runs a synchronous rootmulti.Store.Restore() snapshot restoration which will be fed snapshot chunks until it is complete.

Tendermint will then start fetching and buffering chunks, providing them in order via ABCI ApplySnapshotChunk calls. These dispatch to Manager.RestoreChunk(), which passes the chunks to the ongoing restore process, checking if errors have been encountered yet (e.g. due to checksum mismatches or invalid IAVL data). Once the final chunk is passed, Manager.RestoreChunk() will wait for the restore process to complete before returning.

Once the restore is completed, Tendermint will go on to call the Info ABCI call to fetch the app hash, and compare this against the trusted chain app hash at the snapshot height to verify the restored state. If it matches, Tendermint goes on to process blocks.