This change fixes a bounty by the Juno team. Juno's invariant checks took 10 hours during their most recent chain halt. This PR cuts that down to 30 seconds. See https://github.com/CosmosContracts/bounties#improve-speed-of-invariant-checks. The root problem is deep in the `can-withdraw` invariant check, which calls this repeatedly: https://github.com/cosmos/cosmos-sdk/blob/main/x/distribution/keeper/store.go#L337. Iterators have a chain of parents and in this case creates an iterator from the `cachekv` store. For the genesis file, it has a cache of 500,000+ unsorted entries, which are sorted as strings here: https://github.com/cosmos/cosmos-sdk/blob/main/store/cachekv/store.go#L314. Each delegation from `can-withdraw` uses this cache and many of the cache checks miss or are a very small range. This means very few entries get removed from the unsorted cache and they have to be re-sorted on the next call. With a full cache it takes about 180ms on my machine to sort them. This change introduce a minimum number of entries that will get processed and removed from the unsorted list. It's set at the same value that directs the code to sort them in the first place. This ensures the unsorted values get removed in a relative short amount of time, and amortizes the cost to ensure an individual check does not have to process the entire cache. ## Benchmarks On running the benchmarks included in this change produces: ```shell name old time/op new time/op delta LargeUnsortedMisses-32 21.2s ± 9% 0.0s ± 1% -99.91% (p=0.000 n=20+17) name old alloc/op new alloc/op delta LargeUnsortedMisses-32 1.64GB ± 0% 0.00GB ± 0% -99.83% (p=0.000 n=19+19) name old allocs/op new allocs/op delta LargeUnsortedMisses-32 20.0k ± 0% 41.1k ± 0% +105.23% (p=0.000 n=19+20) ``` ## Invariant checks results This is what the invariant checks for Juno look like with this change (on a Hetzner AX101): ```shell INF starting node with ABCI Tendermint in-process 4:11PM INF Starting multiAppConn service impl=multiAppConn module=proxy 4:11PM INF Starting localClient service connection=query impl=localClient module=abci-client 4:11PM INF Starting localClient service connection=snapshot impl=localClient module=abci-client 4:11PM INF Starting localClient service connection=mempool impl=localClient module=abci-client 4:11PM INF Starting localClient service connection=consensus impl=localClient module=abci-client 4:11PM INF Starting EventBus service impl=EventBus module=events 4:11PM INF Starting PubSub service impl=PubSub module=pubsub 4:11PM INF Starting IndexerService service impl=IndexerService module=txindex 4:11PM INF ABCI Handshake App Info hash= height=0 module=consensus protocol-version=0 software-version=v9.0.0-36-g8fd6f16 4:11PM INF ABCI Replay Blocks appHeight=0 module=consensus stateHeight=0 storeHeight=0 4:12PM INF asserting crisis invariants inv=1/11 module=x/crisis name=gov/module-account 4:12PM INF asserting crisis invariants inv=2/11 module=x/crisis name=distribution/nonnegative-outstanding 4:12PM INF asserting crisis invariants inv=3/11 module=x/crisis name=distribution/can-withdraw 4:12PM INF asserting crisis invariants inv=4/11 module=x/crisis name=distribution/reference-count 4:12PM INF asserting crisis invariants inv=5/11 module=x/crisis name=distribution/module-account 4:12PM INF asserting crisis invariants inv=6/11 module=x/crisis name=bank/nonnegative-outstanding 4:12PM INF asserting crisis invariants inv=7/11 module=x/crisis name=bank/total-supply 4:12PM INF asserting crisis invariants inv=8/11 module=x/crisis name=staking/module-accounts 4:12PM INF asserting crisis invariants inv=9/11 module=x/crisis name=staking/nonnegative-power 4:12PM INF asserting crisis invariants inv=10/11 module=x/crisis name=staking/positive-delegation 4:12PM INF asserting crisis invariants inv=11/11 module=x/crisis name=staking/delegator-shares 4:12PM INF asserted all invariants duration=28383.559601 height=4136532 module=x/crisis ``` ## Alternatives There is another PR which fixes this problem for the Juno genesis file https://github.com/cosmos/cosmos-sdk/pull/12886. However, because of its concurrent nature, it happens to hit a large range relatively early, clearing the unsorted entries and allowing the rest of the checks to not sort it. |
||
|---|---|---|
| .. | ||
| cache | ||
| cachekv | ||
| cachemulti | ||
| dbadapter | ||
| gaskv | ||
| iavl | ||
| internal | ||
| listenkv | ||
| mem | ||
| prefix | ||
| rootmulti | ||
| streaming | ||
| tools/ics23 | ||
| tracekv | ||
| transient | ||
| types | ||
| v2alpha1 | ||
| firstlast.go | ||
| README.md | ||
| reexport.go | ||
| store.go | ||
Store
CacheKV
cachekv.Store is a wrapper KVStore which provides buffered writing / cached reading functionalities over the underlying KVStore.
type Store struct {
cache map[string]cValue
parent types.KVStore
}
Get
Store.Get() checks Store.cache first in order to find if there is any cached value associated with the key. If the value exists, the function returns it. If not, the function calls Store.parent.Get(), sets the key-value pair in the Store.cache, and returns it.
Set
Store.Set() sets the key-value pair to the Store.cache. cValue has the field dirty bool which indicates whether the cached value is different from the underlying value. When Store.Set() cache new pair, the cValue.dirty is set true so when Store.Write() is called it can be written to the underlying store.
Iterator
Store.Iterator() have to traverse on both caches items and the original items. In Store.iterator(), two iterators are generated for each of them, and merged. memIterator is essentially a slice of the KVPairs, used for cached items. mergeIterator is a combination of two iterators, where traverse happens ordered on both iterators.
CacheMulti
cachemulti.Store is a wrapper MultiStore which provides buffered writing / cached reading functionalities over the underlying MutliStore
type Store struct {
db types.CacheKVStore
stores map[types.StoreKey] types.CacheWrap
}
cachemulti.Store branches all substores in its constructor and hold them in Store.stores. Store.GetKVStore() returns the store from Store.stores, and Store.Write() recursively calls CacheWrap.Write() on the substores.
DBAdapter
dbadapter.Store is a adapter for dbm.DB making it fulfilling the KVStore interface.
type Store struct {
dbm.DB
}
dbadapter.Store embeds dbm.DB, so most of the KVStore interface functions are implemented. The other functions(mostly miscellaneous) are manually implemented.
IAVL
iavl.Store is a base-layer self-balancing merkle tree. It is guaranteed that
- Get & set operations are
O(log n), wherenis the number of elements in the tree - Iteration efficiently returns the sorted elements within the range
- Each tree version is immutable and can be retrieved even after a commit(depending on the pruning settings)
Specification and implementation of IAVL tree can be found in https://github.com/cosmos/iavl.
GasKV
gaskv.Store is a wrapper KVStore which provides gas consuming functionalities over the underlying KVStore.
type Store struct {
gasMeter types.GasMeter
gasConfig types.GasConfig
parent types.KVStore
}
When each KVStore methods are called, gaskv.Store automatically consumes appropriate amount of gas depending on the Store.gasConfig.
Prefix
prefix.Store is a wrapper KVStore which provides automatic key-prefixing functionalities over the underlying KVStore.
type Store struct {
parent types.KVStore
prefix []byte
}
When Store.{Get, Set}() is called, the store forwards the call to its parent, with the key prefixed with the Store.prefix.
When Store.Iterator() is called, it does not simply prefix the Store.prefix, since it does not work as intended. In that case, some of the elements are traversed even they are not starting with the prefix.
RootMulti
rootmulti.Store is a base-layer MultiStore where multiple KVStore can be mounted on it and retrieved via object-capability keys. The keys are memory addresses, so it is impossible to forge the key unless an object is a valid owner(or a receiver) of the key, according to the object capability principles.
TraceKV
tracekv.Store is a wrapper KVStore which provides operation tracing functionalities over the underlying KVStore.
type Store struct {
parent types.KVStore
writer io.Writer
context types.TraceContext
}
When each KVStore methods are called, tracekv.Store automatically logs traceOperation to the Store.writer.
type traceOperation struct {
Operation operation
Key string
Value string
Metadata map[string]interface{}
}
traceOperation.Metadata is filled with Store.context when it is not nil. TraceContext is a map[string]interface{}.
Transient
transient.Store is a base-layer KVStore which is automatically discarded at the end of the block.
type Store struct {
dbadapter.Store
}
Store.Store is a dbadapter.Store with a dbm.NewMemDB(). All KVStore methods are reused. When Store.Commit() is called, new dbadapter.Store is assigned, discarding previous reference and making it garbage collected.