tendermint/consensus/replay.go
Anton Kaliaev f6539737de
new pubsub package
comment out failing consensus tests for now

rewrite rpc httpclient to use new pubsub package

import pubsub as tmpubsub, query as tmquery

make event IDs constants
EventKey -> EventTypeKey

rename EventsPubsub to PubSub

mempool does not use pubsub

rename eventsSub to pubsub

new subscribe API

fix channel size issues and consensus tests bugs

refactor rpc client

add missing discardFromChan method

add mutex

rename pubsub to eventBus

remove IsRunning from WSRPCConnection interface (not needed)

add a comment in broadcastNewRoundStepsAndVotes

rename registerEventCallbacks to broadcastNewRoundStepsAndVotes

See https://dave.cheney.net/2014/03/19/channel-axioms

stop eventBuses after reactor tests

remove unnecessary Unsubscribe

return subscribe helper function

move discardFromChan to where it is used

subscribe now returns an err

this gives us ability to refuse to subscribe if pubsub is at its max
capacity.

use context for control overflow

cache queries

handle err when subscribing in replay_test

rename testClientID to testSubscriber

extract var

set channel buffer capacity to 1 in replay_file

fix byzantine_test

unsubscribe from single event, not all events

refactor httpclient to return events to appropriate channels

return failing testReplayCrashBeforeWriteVote test

fix TestValidatorSetChanges

refactor code a bit

fix testReplayCrashBeforeWriteVote

add comment

fix TestValidatorSetChanges

fixes from Bucky's review

update comment [ci skip]

test TxEventBuffer

update changelog

fix TestValidatorSetChanges (2nd attempt)

only do wg.Done when no errors

benchmark event bus

create pubsub server inside NewEventBus

only expose config params (later if needed)

set buffer capacity to 0 so we are not testing cache

new tx event format: key = "Tx" plus a tag {"tx.hash": XYZ}

This should allow to subscribe to all transactions! or a specific one
using a query: "tm.events.type = Tx and tx.hash = '013ABF99434...'"

use TimeoutCommit instead of afterPublishEventNewBlockTimeout

TimeoutCommit is the time a node waits after committing a block, before
it goes into the next height. So it will finish everything from the last
block, but then wait a bit. The idea is this gives it time to hear more
votes from other validators, to strengthen the commit it includes in the
next block. But it also gives it time to hear about new transactions.

waitForBlockWithUpdatedVals

rewrite WAL crash tests

Task:
test that we can recover from any WAL crash.

Solution:
the old tests were relying on event hub being run in the same thread (we
were injecting the private validator's last signature).

when considering a rewrite, we considered two possible solutions: write
a "fuzzy" testing system where WAL is crashing upon receiving a new
message, or inject failures and trigger them in tests using something
like https://github.com/coreos/gofail.

remove sleep

no cs.Lock around wal.Save

test different cases (empty block, non-empty block, ...)

comments

add comments

test 4 cases: empty block, non-empty block, non-empty block with smaller part size, many blocks

fixes as per Bucky's last review

reset subscriptions on UnsubscribeAll

use a simple counter to track message for which we panicked

also, set a smaller part size for all test cases
2017-10-30 00:32:22 -05:00

396 lines
13 KiB
Go

package consensus
import (
"bytes"
"errors"
"fmt"
"hash/crc32"
"io"
"reflect"
"strconv"
"strings"
"time"
abci "github.com/tendermint/abci/types"
auto "github.com/tendermint/tmlibs/autofile"
cmn "github.com/tendermint/tmlibs/common"
"github.com/tendermint/tmlibs/log"
"github.com/tendermint/tendermint/proxy"
sm "github.com/tendermint/tendermint/state"
"github.com/tendermint/tendermint/types"
"github.com/tendermint/tendermint/version"
)
var crc32c = crc32.MakeTable(crc32.Castagnoli)
// Functionality to replay blocks and messages on recovery from a crash.
// There are two general failure scenarios: failure during consensus, and failure while applying the block.
// The former is handled by the WAL, the latter by the proxyApp Handshake on restart,
// which ultimately hands off the work to the WAL.
//-----------------------------------------
// recover from failure during consensus
// by replaying messages from the WAL
// Unmarshal and apply a single message to the consensus state
// as if it were received in receiveRoutine
// Lines that start with "#" are ignored.
// NOTE: receiveRoutine should not be running
func (cs *ConsensusState) readReplayMessage(msg *TimedWALMessage, newStepCh chan interface{}) error {
// skip meta messages
if _, ok := msg.Msg.(EndHeightMessage); ok {
return nil
}
// for logging
switch m := msg.Msg.(type) {
case types.EventDataRoundState:
cs.Logger.Info("Replay: New Step", "height", m.Height, "round", m.Round, "step", m.Step)
// these are playback checks
ticker := time.After(time.Second * 2)
if newStepCh != nil {
select {
case mi := <-newStepCh:
m2 := mi.(types.EventDataRoundState)
if m.Height != m2.Height || m.Round != m2.Round || m.Step != m2.Step {
return fmt.Errorf("RoundState mismatch. Got %v; Expected %v", m2, m)
}
case <-ticker:
return fmt.Errorf("Failed to read off newStepCh")
}
}
case msgInfo:
peerKey := m.PeerKey
if peerKey == "" {
peerKey = "local"
}
switch msg := m.Msg.(type) {
case *ProposalMessage:
p := msg.Proposal
cs.Logger.Info("Replay: Proposal", "height", p.Height, "round", p.Round, "header",
p.BlockPartsHeader, "pol", p.POLRound, "peer", peerKey)
case *BlockPartMessage:
cs.Logger.Info("Replay: BlockPart", "height", msg.Height, "round", msg.Round, "peer", peerKey)
case *VoteMessage:
v := msg.Vote
cs.Logger.Info("Replay: Vote", "height", v.Height, "round", v.Round, "type", v.Type,
"blockID", v.BlockID, "peer", peerKey)
}
cs.handleMsg(m)
case timeoutInfo:
cs.Logger.Info("Replay: Timeout", "height", m.Height, "round", m.Round, "step", m.Step, "dur", m.Duration)
cs.handleTimeout(m, cs.RoundState)
default:
return fmt.Errorf("Replay: Unknown TimedWALMessage type: %v", reflect.TypeOf(msg.Msg))
}
return nil
}
// replay only those messages since the last block.
// timeoutRoutine should run concurrently to read off tickChan
func (cs *ConsensusState) catchupReplay(csHeight int) error {
// set replayMode
cs.replayMode = true
defer func() { cs.replayMode = false }()
// Ensure that ENDHEIGHT for this height doesn't exist
// NOTE: This is just a sanity check. As far as we know things work fine without it,
// and Handshake could reuse ConsensusState if it weren't for this check (since we can crash after writing ENDHEIGHT).
gr, found, err := cs.wal.SearchForEndHeight(uint64(csHeight))
if gr != nil {
gr.Close()
}
if found {
return fmt.Errorf("WAL should not contain #ENDHEIGHT %d.", csHeight)
}
// Search for last height marker
gr, found, err = cs.wal.SearchForEndHeight(uint64(csHeight - 1))
if err == io.EOF {
cs.Logger.Error("Replay: wal.group.Search returned EOF", "#ENDHEIGHT", csHeight-1)
} else if err != nil {
return err
}
if !found {
return errors.New(cmn.Fmt("Cannot replay height %d. WAL does not contain #ENDHEIGHT for %d.", csHeight, csHeight-1))
}
defer gr.Close()
cs.Logger.Info("Catchup by replaying consensus messages", "height", csHeight)
var msg *TimedWALMessage
dec := WALDecoder{gr}
for {
msg, err = dec.Decode()
if err == io.EOF {
break
} else if err != nil {
return err
}
// NOTE: since the priv key is set when the msgs are received
// it will attempt to eg double sign but we can just ignore it
// since the votes will be replayed and we'll get to the next step
if err := cs.readReplayMessage(msg, nil); err != nil {
return err
}
}
cs.Logger.Info("Replay: Done")
return nil
}
//--------------------------------------------------------------------------------
// Parses marker lines of the form:
// #ENDHEIGHT: 12345
func makeHeightSearchFunc(height int) auto.SearchFunc {
return func(line string) (int, error) {
line = strings.TrimRight(line, "\n")
parts := strings.Split(line, " ")
if len(parts) != 2 {
return -1, errors.New("Line did not have 2 parts")
}
i, err := strconv.Atoi(parts[1])
if err != nil {
return -1, errors.New("Failed to parse INFO: " + err.Error())
}
if height < i {
return 1, nil
} else if height == i {
return 0, nil
} else {
return -1, nil
}
}
}
//----------------------------------------------
// Recover from failure during block processing
// by handshaking with the app to figure out where
// we were last and using the WAL to recover there
type Handshaker struct {
state *sm.State
store types.BlockStore
logger log.Logger
nBlocks int // number of blocks applied to the state
}
func NewHandshaker(state *sm.State, store types.BlockStore) *Handshaker {
return &Handshaker{state, store, log.NewNopLogger(), 0}
}
func (h *Handshaker) SetLogger(l log.Logger) {
h.logger = l
}
func (h *Handshaker) NBlocks() int {
return h.nBlocks
}
// TODO: retry the handshake/replay if it fails ?
func (h *Handshaker) Handshake(proxyApp proxy.AppConns) error {
// handshake is done via info request on the query conn
res, err := proxyApp.Query().InfoSync(abci.RequestInfo{version.Version})
if err != nil {
return errors.New(cmn.Fmt("Error calling Info: %v", err))
}
blockHeight := int(res.LastBlockHeight) // XXX: beware overflow
appHash := res.LastBlockAppHash
h.logger.Info("ABCI Handshake", "appHeight", blockHeight, "appHash", fmt.Sprintf("%X", appHash))
// TODO: check version
// replay blocks up to the latest in the blockstore
_, err = h.ReplayBlocks(appHash, blockHeight, proxyApp)
if err != nil {
return errors.New(cmn.Fmt("Error on replay: %v", err))
}
h.logger.Info("Completed ABCI Handshake - Tendermint and App are synced", "appHeight", blockHeight, "appHash", fmt.Sprintf("%X", appHash))
// TODO: (on restart) replay mempool
return nil
}
// Replay all blocks since appBlockHeight and ensure the result matches the current state.
// Returns the final AppHash or an error
func (h *Handshaker) ReplayBlocks(appHash []byte, appBlockHeight int, proxyApp proxy.AppConns) ([]byte, error) {
storeBlockHeight := h.store.Height()
stateBlockHeight := h.state.LastBlockHeight
h.logger.Info("ABCI Replay Blocks", "appHeight", appBlockHeight, "storeHeight", storeBlockHeight, "stateHeight", stateBlockHeight)
// If appBlockHeight == 0 it means that we are at genesis and hence should send InitChain
if appBlockHeight == 0 {
validators := types.TM2PB.Validators(h.state.Validators)
proxyApp.Consensus().InitChainSync(abci.RequestInitChain{validators})
}
// First handle edge cases and constraints on the storeBlockHeight
if storeBlockHeight == 0 {
return appHash, h.checkAppHash(appHash)
} else if storeBlockHeight < appBlockHeight {
// the app should never be ahead of the store (but this is under app's control)
return appHash, sm.ErrAppBlockHeightTooHigh{storeBlockHeight, appBlockHeight}
} else if storeBlockHeight < stateBlockHeight {
// the state should never be ahead of the store (this is under tendermint's control)
cmn.PanicSanity(cmn.Fmt("StateBlockHeight (%d) > StoreBlockHeight (%d)", stateBlockHeight, storeBlockHeight))
} else if storeBlockHeight > stateBlockHeight+1 {
// store should be at most one ahead of the state (this is under tendermint's control)
cmn.PanicSanity(cmn.Fmt("StoreBlockHeight (%d) > StateBlockHeight + 1 (%d)", storeBlockHeight, stateBlockHeight+1))
}
// Now either store is equal to state, or one ahead.
// For each, consider all cases of where the app could be, given app <= store
if storeBlockHeight == stateBlockHeight {
// Tendermint ran Commit and saved the state.
// Either the app is asking for replay, or we're all synced up.
if appBlockHeight < storeBlockHeight {
// the app is behind, so replay blocks, but no need to go through WAL (state is already synced to store)
return h.replayBlocks(proxyApp, appBlockHeight, storeBlockHeight, false)
} else if appBlockHeight == storeBlockHeight {
// We're good!
return appHash, h.checkAppHash(appHash)
}
} else if storeBlockHeight == stateBlockHeight+1 {
// We saved the block in the store but haven't updated the state,
// so we'll need to replay a block using the WAL.
if appBlockHeight < stateBlockHeight {
// the app is further behind than it should be, so replay blocks
// but leave the last block to go through the WAL
return h.replayBlocks(proxyApp, appBlockHeight, storeBlockHeight, true)
} else if appBlockHeight == stateBlockHeight {
// We haven't run Commit (both the state and app are one block behind),
// so replayBlock with the real app.
// NOTE: We could instead use the cs.WAL on cs.Start,
// but we'd have to allow the WAL to replay a block that wrote it's ENDHEIGHT
h.logger.Info("Replay last block using real app")
return h.replayBlock(storeBlockHeight, proxyApp.Consensus())
} else if appBlockHeight == storeBlockHeight {
// We ran Commit, but didn't save the state, so replayBlock with mock app
abciResponses := h.state.LoadABCIResponses()
mockApp := newMockProxyApp(appHash, abciResponses)
h.logger.Info("Replay last block using mock app")
return h.replayBlock(storeBlockHeight, mockApp)
}
}
cmn.PanicSanity("Should never happen")
return nil, nil
}
func (h *Handshaker) replayBlocks(proxyApp proxy.AppConns, appBlockHeight, storeBlockHeight int, mutateState bool) ([]byte, error) {
// App is further behind than it should be, so we need to replay blocks.
// We replay all blocks from appBlockHeight+1.
//
// Note that we don't have an old version of the state,
// so we by-pass state validation/mutation using sm.ExecCommitBlock.
// This also means we won't be saving validator sets if they change during this period.
//
// If mutateState == true, the final block is replayed with h.replayBlock()
var appHash []byte
var err error
finalBlock := storeBlockHeight
if mutateState {
finalBlock -= 1
}
for i := appBlockHeight + 1; i <= finalBlock; i++ {
h.logger.Info("Applying block", "height", i)
block := h.store.LoadBlock(i)
appHash, err = sm.ExecCommitBlock(proxyApp.Consensus(), block, h.logger)
if err != nil {
return nil, err
}
h.nBlocks += 1
}
if mutateState {
// sync the final block
return h.replayBlock(storeBlockHeight, proxyApp.Consensus())
}
return appHash, h.checkAppHash(appHash)
}
// ApplyBlock on the proxyApp with the last block.
func (h *Handshaker) replayBlock(height int, proxyApp proxy.AppConnConsensus) ([]byte, error) {
mempool := types.MockMempool{}
block := h.store.LoadBlock(height)
meta := h.store.LoadBlockMeta(height)
if err := h.state.ApplyBlock(types.NopEventBus{}, proxyApp, block, meta.BlockID.PartsHeader, mempool); err != nil {
return nil, err
}
h.nBlocks += 1
return h.state.AppHash, nil
}
func (h *Handshaker) checkAppHash(appHash []byte) error {
if !bytes.Equal(h.state.AppHash, appHash) {
panic(errors.New(cmn.Fmt("Tendermint state.AppHash does not match AppHash after replay. Got %X, expected %X", appHash, h.state.AppHash)).Error())
return nil
}
return nil
}
//--------------------------------------------------------------------------------
// mockProxyApp uses ABCIResponses to give the right results
// Useful because we don't want to call Commit() twice for the same block on the real app.
func newMockProxyApp(appHash []byte, abciResponses *sm.ABCIResponses) proxy.AppConnConsensus {
clientCreator := proxy.NewLocalClientCreator(&mockProxyApp{
appHash: appHash,
abciResponses: abciResponses,
})
cli, _ := clientCreator.NewABCIClient()
cli.Start()
return proxy.NewAppConnConsensus(cli)
}
type mockProxyApp struct {
abci.BaseApplication
appHash []byte
txCount int
abciResponses *sm.ABCIResponses
}
func (mock *mockProxyApp) DeliverTx(tx []byte) abci.Result {
r := mock.abciResponses.DeliverTx[mock.txCount]
mock.txCount += 1
return abci.Result{
r.Code,
r.Data,
r.Log,
}
}
func (mock *mockProxyApp) EndBlock(height uint64) abci.ResponseEndBlock {
mock.txCount = 0
return mock.abciResponses.EndBlock
}
func (mock *mockProxyApp) Commit() abci.Result {
return abci.NewResultOK(mock.appHash, "")
}