cerc/go-ethereum
statediff functionality packaged as a plugeth plugin.
This package provides a PluGeth plugin implementing an
auxiliary service that asynchronously computes changes in the go-ethereum
state trie. The service
continuously listens for chain updates and builds state diffs, then relays the data to RPC
subscribers or writes them directly to Postgres as IPLD objects.
It also exposes RPC endpoints for fetching or writing to Postgres the state diff at a specific block height or for a specific block hash. This operates on historical block and state data, and so depends on a complete state archive.
Data is emitted in this differential format in order to make it feasible to the entire Ethereum state and publish it to IPLD (including intermediate state and storage trie nodes). If this service is run continuously from genesis, the entire state at any block can be materialized from the cumulative differentials up to that point.
The primary interface type is Payload
, which serves as the main interface for accessing data in a
service. It packages various data components such as block RLP, total difficulty, receipts RLP, and
state object RLP. This encapsulates all of the differential data at a given block, and allows us to
index the entire Ethereum data structure as hash-linked IPLD objects.
The StateObject
type represents the final diff output structure, including an array of state leaf
nodes and IPLD objects. For convenience, we also associate this object with the block number and
hash.
State leaf nodes contain information about account changes, including whether they are removed, an account wrapper with account details and identifiers, and an array of storage leaf nodes representing storage changes. The IPLD type encapsulates CID-content pairs, used for code mappings and trie node (both intermediate and leaf) IPLD objects.
// Payload packages the data to send to state diff subscriptions
type Payload struct {
BlockRlp []byte `json:"blockRlp"`
TotalDifficulty *big.Int `json:"totalDifficulty"`
ReceiptsRlp []byte `json:"receiptsRlp"`
StateObjectRlp []byte `json:"stateObjectRlp" gencodec:"required"`
// ...
}
// in package "types":
// StateObject is the final output structure from the builder
type StateObject struct {
BlockNumber *big.Int `json:"blockNumber" gencodec:"required"`
BlockHash common.Hash `json:"blockHash" gencodec:"required"`
Nodes []StateLeafNode `json:"nodes" gencodec:"required"`
IPLDs []IPLD `json:"iplds"`
}
// StateLeafNode holds the data for a single state diff leaf node
type StateLeafNode struct {
Removed bool
AccountWrapper AccountWrapper
StorageDiff []StorageLeafNode
}
// AccountWrapper is used to temporarily associate the unpacked node with its raw values
type AccountWrapper struct {
Account *types.StateAccount
LeafKey []byte
CID string
}
// StorageLeafNode holds the data for a single storage diff node leaf node
type StorageLeafNode struct {
Removed bool
Value []byte
LeafKey []byte
CID string
}
// IPLD holds a cid:content pair, e.g. for codehash to code mappings or for intermediate node IPLD objects
type IPLD struct {
CID string
Content []byte
}
The service is started when the plugin library is loaded by PluGeth and runs as an auxiliary component of the node as it syncs.
This service introduces a CLI flag namespace statediff
. Note that PluGeth plugin arguments must be separated from geth arguments by --
, e.g. geth --datadir data -- --statediff
.
--statediff
is used to enable the service--statediff.writing
is used to tell the service to write state diff objects it produces from syncedChainEvent
s directly to a configured Postgres database--statediff.workers
is used to set the number of concurrent workers to process state diff objects and write them into the database--statediff.db.type
is the type of database we write out to (current options:postgres
,dump
,file
)--statediff.dump.dst
is the destination to write to when operating in database dump mode (stdout
,stderr
,discard
)--statediff.db.driver
is the specific driver to use for the database (current options for postgres:pgx
andsqlx
)--statediff.db.host
is the hostname address to dial to connect to the database--statediff.db.port
is the port to dial to connect to the database--statediff.db.name
is the name of the database to connect to--statediff.db.user
is the user to connect to the database as--statediff.db.password
is the password to use to connect to the database--statediff.db.conntimeout
is the connection timeout (in seconds)--statediff.db.maxconns
is the maximum number of database connections--statediff.db.minconns
is the minimum number of database connections--statediff.db.maxidleconns
is the maximum number of idle connections--statediff.db.maxconnidletime
is the maximum lifetime for an idle connection (in seconds)--statediff.db.maxconnlifetime
is the maximum lifetime for a connection (in seconds)--statediff.db.nodeid
is the node id to use in the Postgres database--statediff.db.clientname
is the client name to use in the Postgres database--statediff.db.upsert
whether or not the service, when operating in a direct database writing mode, should overwrite any existing conflicting data--statediff.file.path
full path (including filename) to write statediff data out to when operating in file mode--statediff.file.wapath
full path (including filename) to write statediff watched addresses out to when operating in file mode
The service can only operate in full sync mode (--syncmode=full
), but only the historical RPC endpoints require an archive node (--gcmode=archive
)
e.g.
geth --syncmode=full --gcmode=archive -- --statediff --statediff.writing --statediff.db.type=postgres --statediff.db.driver=sqlx --statediff.db.host=localhost --statediff.db.port=5432 --statediff.db.name=cerc_testing --statediff.db.user=postgres --statediff.db.nodeid=nodeid --statediff.db.clientname=clientname
When operating in --statediff.db.type=file
mode, the service will save SQL statements to the file
specified by --statediff.file.path
. It's important to note that these SQL statements are written
without any ON CONFLICT
constraint checks. This omission allows us to:
- horizontally expand the production of SQL statements,
- merge the individual SQL files generated,
- remove duplicates using Unix tools (
sort statediff.sql | uniq
orsort -u statediff.sql
), - perform bulk loading using psql (
psql db_name --set ON_ERROR_STOP=on -f statediff.sql
), - and then reinstate our primary and foreign key constraints and indexes.
The state diffing service exposes both a websocket subscription endpoint, and a number of HTTP unary endpoints for retrieving data payloads.
Each of these endpoints requires a set of parameters provided by the caller:
// Params is used to carry in parameters from subscribing/requesting clients configuration
type Params struct {
IntermediateStateNodes bool
IntermediateStorageNodes bool
IncludeBlock bool
IncludeReceipts bool
IncludeTD bool
IncludeCode bool
WatchedAddresses []common.Address
}
Using these params we can tell the service:
- whether to include state and/or storage intermediate nodes
- whether to include the associated block (header, uncles, and transactions)
- whether to include the associated receipts
- whether to include the total difficulty for this block
- whether to include the set of code hashes and code for contracts deployed in this block, and
- whether to limit the diffing process to a list of specific addresses.
A websocket-supporting RPC endpoint is exposed for subscribing to state diff StateObjects
that come off the head of the chain while the geth node syncs.
// Stream is a subscription endpoint that fires off state diff payloads as they are created
Stream(ctx context.Context, params Params) (*rpc.Subscription, error)
To expose this endpoint the node needs to have the websocket server turned on (--ws
),
and the statediff
namespace exposed (--ws.api=statediff
).
Go code subscriptions to this endpoint can be created using the rpc.Client.Subscribe()
method,
with the "statediff" namespace, a statediff.Payload
channel, and the name of the statediff api's rpc method: "stream".
e.g.
cli, err := rpc.Dial("ipcPathOrWsURL")
if err != nil {
// handle error
}
stateDiffPayloadChan := make(chan statediff.Payload, 20000)
methodName := "stream"
params := statediff.Params{
IncludeBlock: true,
IncludeTD: true,
IncludeReceipts: true,
IntermediateStorageNodes: true,
IntermediateStateNodes: true,
}
rpcSub, err := cli.Subscribe(context.Background(), statediff.APIName, stateDiffPayloadChan, methodName, params)
if err != nil {
// handle error
}
for {
select {
case stateDiffPayload := <- stateDiffPayloadChan:
// process the payload
case err := <- rpcSub.Err():
// handle rpc subscription error
}
}
The service also exposes unary RPC endpoints for retrieving the state diff StateObject
for a specific block height/hash.
// StateDiffAt returns a state diff payload at the specific blockheight
StateDiffAt(ctx context.Context, blockNumber uint64, params Params) (*Payload, error)
// StateDiffFor returns a state diff payload for the specific blockhash
StateDiffFor(ctx context.Context, blockHash common.Hash, params Params) (*Payload, error)
To expose this endpoint the node needs to have the HTTP server turned on (--http
),
and the statediff
namespace exposed (--http.api=statediff
).
If --statediff.writing
is enabled, the service will convert the StateObject
s and all associated
data into IPLD objects, persist them directly to Postgres, and generate secondary indexes around the
IPLD data.
The schema and migrations for this Postgres database are defined in https://github.com/cerc-io/ipld-eth-db.
If enabled, direct indexing will be triggered on every ChainEvent
, writing diffs for all new
blocks as they are received. However, the service also provides methods for clients to trigger and
track this process:
- The
WriteStateDiffAt
method directly writes a state diff object to the database at a specific block height. - Likewise, the
WriteStateDiffFor
method directly writes a state diff object to the database for a specific block hash - The
StreamWrites
method sets up a subscription to stream the status of completed calls to the above methods. - The
WatchAddress
method enables the modification of the watched addresses list, restricting direct indexing for a given operation and arguments.
Our Postgres schemas are built around a single IPFS backing Postgres IPLD blockstore table
(ipld.blocks
) that conforms with
go-ds-sql. All IPLD objects
are stored in this table, where key
is the CID for the IPLD object and data
contains the bytes
for the IPLD block (in the case of all Ethereum IPLDs, this is the RLP byte encoding of the Ethereum
object).
The IPLD objects in this table can be traversed using an IPLD DAG interface, but since this table only maps CID to raw IPLD object it is not very suitable for looking up Ethereum objects by their constituent fields (e.g. by tx source/recipient, state/storage trie path). To improve the accessibility of these objects we create an Ethereum advanced data layout (ADL) by generating secondary indexes on top of the raw IPLDs in other Postgres tables.
These secondary index tables fall under the eth
schema and follow an {objectType}_cids
naming
convention. These tables provide a view into individual fields of the underlying Ethereum IPLD
objects, allowing lookups on these fields, and reference the raw IPLD objects stored in
ipld.blocks
by CID. Additionally, these tables maintain the hash-linked nature of Ethereum
objects to one another, e.g. a storage trie node entry in the storage_cids
table contains a
state_leaf_key
field referencing the state_cids
entry for the state trie node of its owning
contract, and that state_cids
entry in turn contains a header_id
field referencing the
block_hash
of the header_cids
entry for the block in which these state and storage nodes were
updated (diffed).
On mainnet this process is extremely IO intensive and requires significant resources to allow it to keep up with the head of the chain. The state diff processing time for a specific block is dependent on the number and complexity of the state changes that occur in a block and the number of updated state nodes that are available in the in-memory cache vs must be retrieved from disc.
If memory permits, one means of improving the efficiency of this process is to increase the
in-memory trie cache allocation. This can be done by increasing the overall --cache
allocation
and/or by increasing the % of the cache allocated to trie usage with --cache.trie
.