tla-plus/RaftMerge/RaftMerge.tla
2018-03-22 14:56:07 +08:00

539 lines
19 KiB
Plaintext

------------------------------ MODULE RaftMerge -------------------------------
\* This is the formal specification for the multi-raft region merge algorithm
\* of TiKV.
\*
\* The whole data is divided into multiple shards called regions, and each
\* region is replicated to several stores comprising a Raft group. Two regions
\* can merge into a larger one if one region is reasonably small.
\*
\* This specification asserts two regions named A and B are replicated to the
\* same set of stores. Each region has leader on store LeaderA and LeaderB
\* respectively.
\*
\* Notice that TiKV uses a slightly different Raft model compared with Ongaro's
\* original Raft implementation. A log is truly committed if the log is applied
\* to the state machine, and then server will return the result to client.
\* commit_index is only a marker, log may be dropped even if commit_index goes
\* beyond that log.
EXTENDS Integers, FiniteSets, Sequences, TLC
CONSTANTS Store, Region
CONSTANTS RegionA, RegionB
CONSTANTS LeaderA, LeaderB
ASSUME /\ Region = {RegionA, RegionB}
/\ LeaderA \in Store
/\ LeaderB \in Store
\* If TRUE, a rollback will be performed in place of merge.
CONSTANTS WillPerformRollback
ASSUME WillPerformRollback \in BOOLEAN
\* The minimum size of servers to be a quorum.
\*
\* The motivation of adding this constant is that, Raft with 1 leader and 2
\* followers is rather slow for TLC to verify (the space to explore is fricking
\* big). We relax Raft with only 1 leader and 1 follower. By setting quorum size
\* to 1 as a hack, we can simulate some states that a follower can have some
\* logs lag behind, a scenario which can happen when there are at least 3
\* servers in a Raft group. This setting does not affect the correctness of Raft
\* protocol because we assume there is no leader switch.
\*
\* If set to 0, a set of servers is considered as a quorum if it contains at
\* least half of servers.
CONSTANTS QuorumSize
\* Maximum number of client requests.
CONSTANTS MaxClientRequests
\* The current index of client request, should be smaller than
\* MaxClientRequests.
VARIABLES client_requests_index
\* Message types for Raft log synchronization.
CONSTANTS AppendEntriesRequest, AppendEntriesReply
\* A set of messages representing RPC requests and responses sent from one
\* server to another.
VARIABLES messages
\* The data structures in C. MAXS = |Store|.
\*
\* enum Log { LogNormal, LogPreMerge, LogMerge, LogRollback };
\*
\* enum RegionState { RegionNormal, RegionTombStone, RegionMerging };
\*
\* struct Raft {
\* bool is_leader;
\* vector<Log> logs;
\* int commit_index;
\* int apply_index;
\* int num_applied; // number of applied normal logs
\* int match_index[MAXS]; // leader only
\* };
\*
\* struct Store {
\* Raft raft[2]; // 2 for two regions
\* RegionState region[2]; // 2 for two regions
\* } stores[MAXS];
\*
\* Note for ease of implementation, we use two 2-dimension arrays raft[MAXS][2]
\* and region[MAXS][2].
\*
\* Also note that different from a real-world implementation, we don't
\* introduce the concept of `epoch` here, which is used to figure out whether
\* the configuration of one region has changed. Epoch matters when we are
\* applying the logs into state machine, if it is stale, we will skip all
\* later non-admin logs. Epoch will be changed when we are applying admin logs.
\* Log types.
\* The logs are divided into two categories, normal logs and admin logs.
\* Logs apart from LogNormal are admin logs.
CONSTANTS LogNormal, \* RegionB only
LogPreMerge,
LogMerge,
LogRollback
\* Region state types.
CONSTANTS RegionNormal,
RegionTombStone,
RegionMerging
VARIABLES raft, region
client_vars == <<client_requests_index>>
vars == <<messages, raft, region, client_vars>>
-------------------------------------------------------------------------------
\* Helper functions.
\* Return true if the server set is a quorum.
IsQuorum(svrs) ==
IF QuorumSize = 0
THEN Cardinality(svrs) * 2 > Cardinality(Store)
ELSE Cardinality(svrs) >= QuorumSize
\* Return the minimum value from a set, or undefined if the set is empty.
Min(s) == CHOOSE x \in s : \A y \in s : x <= y
\* Return the maximum value from a set, or undefined if the set is empty.
Max(s) == CHOOSE x \in s : \A y \in s : x >= y
\* Add a message to the set of messages.
Send(m) == messages' = messages \cup {m}
\* Remove a message from the set of messages. Used when a server is done
\* processing a message.
Discard(m) == messages' = messages \ {m}
\* Combination of Send and Discard.
Reply(reply, request) == messages' = (messages \cup {reply}) \ {request}
-------------------------------------------------------------------------------
\* State transitions and message handlers for Raft.
\* Leader i of region r sends j an AppendEntries request containing 1 entry.
AppendEntries(i, j, r) ==
/\ i /= j
/\ raft[i][r].is_leader
/\ LET
next_index == raft[i][r].match_index[j] + 1
IN
/\ next_index <= Len(raft[i][r].logs)
/\ Send([type |-> AppendEntriesRequest,
region |-> r,
source |-> i,
dest |-> j,
entry |-> raft[i][r].logs[next_index],
entry_index |-> next_index,
commit_index |-> Min({next_index, raft[i][r].commit_index})])
/\ UNCHANGED <<raft, region, client_vars>>
\* Leader i of region r advances its commitIndex.
AdvanceCommitIndex(i, r) ==
/\ raft[i][r].is_leader
/\ LET
\* The set of servers that agree up through index.
Agree(index) == {i} \cup {k \in Store : raft[i][r].match_index[k] >= index}
\* The maximum indexes for which a quorum agrees.
agree_indexes == {index \in 1..Len(raft[i][r].logs) : IsQuorum(Agree(index))}
\* New value for commitIndex'[i].
new_commit_index == Max(agree_indexes \cup {raft[i][r].commit_index})
IN
raft' = [raft EXCEPT ![i][r].commit_index = new_commit_index]
/\ UNCHANGED <<messages, region, client_vars>>
\* Server i of region r receives an AppendEntries request from server j.
HandleAppendEntriesRequest(i, j, r, m) ==
\/ \* Append this log entry if it is new.
/\ m.entry_index = Len(raft[i][r].logs) + 1
/\ LET
new_logs == Append(raft[i][r].logs, m.entry)
new_commit_index == Max({raft[i][r].commit_index, m.commit_index})
IN
/\ raft' = [raft EXCEPT ![i][r].logs = new_logs,
![i][r].commit_index = new_commit_index]
/\ Reply([type |-> AppendEntriesReply,
region |-> r,
source |-> i,
dest |-> j,
match_index |-> Len(new_logs)],
m)
/\ UNCHANGED <<region, client_vars>>
\/ \* We already have this log entry, discarding this message.
/\ m.entry_index <= Len(raft[i][r].logs)
/\ Discard(m)
/\ UNCHANGED <<raft, region, client_vars>>
\* Server i of region r receives an AppendEntries reply from server j.
HandleAppendEntriesReply(i, j, r, m) ==
/\ raft' = [raft EXCEPT ![i][r].match_index = [@ EXCEPT ![j] = Max({@, m.match_index})]]
/\ Discard(m)
/\ UNCHANGED <<region, client_vars>>
\* Poll a message from message set, and invoke the corresponding RPC handler.
Receive(m) ==
LET
i == m.dest
j == m.source
r == m.region
IN
\/ /\ m.type = AppendEntriesRequest
/\ HandleAppendEntriesRequest(i, j, r, m)
\/ /\ m.type = AppendEntriesReply
/\ HandleAppendEntriesReply(i, j, r, m)
\* Leader i of region r receives a client request to append a log.
ClientRequest(i, r, log) ==
/\ raft[i][r].is_leader
/\ region[i][r] = RegionNormal
/\ client_requests_index < MaxClientRequests
/\ LET
new_logs == Append(raft[i][r].logs, log)
new_match_index == [raft[i][r].match_index EXCEPT ![i] = @ + 1]
IN
/\ raft' = [raft EXCEPT ![i][r].logs = new_logs,
![i][r].match_index = new_match_index]
/\ client_requests_index' = client_requests_index + 1
/\ UNCHANGED <<messages, region>>
-------------------------------------------------------------------------------
\* State transitions for Raft merge, and log applying.
\* Internal requests for Raft merge.
\* Assume raft[i][r].is_leader, i.e., only leader can handle internal requests.
InternalRequest(i, r, log) ==
LET
new_logs == Append(raft[i][r].logs, log)
new_match_index == [raft[i][r].match_index EXCEPT ![i] = @ + 1]
IN
[raft EXCEPT ![i][r].logs = new_logs,
![i][r].match_index = new_match_index]
\* Send merge request to the leader of Region B.
ProposeMergeRequest(i) ==
/\ raft[i][RegionB].is_leader
/\ region[i][RegionB] = RegionNormal
/\ \* This request should be sent only once.
Len(SelectSeq(raft[i][RegionB].logs, LAMBDA log : log.type = LogPreMerge)) = 0
/\ raft' = InternalRequest(
i, RegionB,
[type |-> LogPreMerge,
min_index |-> 1 + Min({raft[i][RegionB].match_index[j] : j \in Store})]
)
/\ UNCHANGED <<messages, region, client_vars>>
\* Perform rollback by appending a rollback log on leader B.
PerformRollbackRequest(i) ==
/\ WillPerformRollback = TRUE
/\ raft[i][RegionB].is_leader
/\ region[i][RegionB] = RegionMerging
/\ \* This log should be appended only once.
Len(SelectSeq(raft[i][RegionB].logs, LAMBDA log : log.type = LogRollback)) = 0
/\ raft' = InternalRequest(i, RegionB, [type |-> LogRollback])
/\ UNCHANGED <<messages, region, client_vars>>
\* Return TRUE if there is a log applicable to the state machine.
\* A log is applicable if it is committed, and the target region is not in
\* TombStone state.
LogAppliable(i, r) ==
/\ raft[i][r].apply_index < raft[i][r].commit_index
/\ region[i][r] /= RegionTombStone
\* Apply LogPreMerge.
ApplyPreMergeLog(i) ==
LET
next_index == raft[i][RegionB].apply_index + 1
IN
/\ LogAppliable(i, RegionB)
/\ raft[i][RegionB].logs[next_index].type = LogPreMerge
/\ IF raft[i][RegionA].is_leader
THEN
\* If this store is the leader of regionA, make a merge proposal, and
\* advance apply_index.
LET
min_index == raft[i][RegionB].logs[next_index].min_index
commit_index == next_index
fetch_logs == SubSeq(raft[i][RegionB].logs, min_index, commit_index)
IN
raft' = [InternalRequest(
i, RegionA,
[type |-> LogMerge,
min_index |-> min_index,
commit_index |-> commit_index,
entries |-> fetch_logs]
)
EXCEPT ![i][RegionB].apply_index = next_index]
ELSE
\* Otherwise, only advance apply_index.
raft' = [raft EXCEPT ![i][RegionB].apply_index = next_index]
/\ region' = [region EXCEPT ![i][RegionB] = RegionMerging]
/\ UNCHANGED <<messages, client_vars>>
\* Apply LogMerge.
\*
\* This action is roughly divided into two sub-actions, and executed separately.
\* The first step copies the logs to region B, to ensure it in sync with leader
\* B. The second step waits until the copied logs in the first step are applied,
\* then advances apply_index and marks this region as tombstone.
ApplyMergeLogStep1(i) ==
LET
next_index == raft[i][RegionA].apply_index + 1
min_index == raft[i][RegionA].logs[next_index].min_index
commit_index == raft[i][RegionA].logs[next_index].commit_index
new_logs ==
LET
old_logs == raft[i][RegionB].logs
entries == raft[i][RegionA].logs[next_index].entries
IN
IF commit_index <= Len(raft[i][RegionB].logs)
THEN old_logs
ELSE old_logs \o SubSeq(entries, Len(old_logs) - min_index + 2, Len(entries))
IN
/\ raft' = [raft EXCEPT ![i][RegionB].logs = new_logs,
![i][RegionB].commit_index = Max({@, commit_index})]
/\ UNCHANGED <<messages, region, client_vars>>
ApplyMergeLogStep2(i) ==
LET
next_index == raft[i][RegionA].apply_index + 1
commit_index == raft[i][RegionA].logs[next_index].commit_index
IN
/\ \* Lag logs have been applied.
raft[i][RegionB].apply_index >= commit_index
/\ raft' = [raft EXCEPT ![i][RegionA].apply_index = next_index]
/\ IF WillPerformRollback = FALSE
THEN
region' = [region EXCEPT ![i][RegionB] = RegionTombStone]
ELSE
UNCHANGED <<region>>
/\ UNCHANGED <<messages, client_vars>>
ApplyMergeLog(i) ==
LET
next_index == raft[i][RegionA].apply_index + 1
IN
/\ LogAppliable(i, RegionA)
/\ raft[i][RegionA].logs[next_index].type = LogMerge
/\ \/ ApplyMergeLogStep1(i)
\/ ApplyMergeLogStep2(i)
\* Apply LogRollback.
\* As we skip all logs after PreMerge log, rollback just marks the state of
\* region as normal.
ApplyRollbackLog(i) ==
LET
next_index == raft[i][RegionB].apply_index + 1
IN
/\ LogAppliable(i, RegionB)
/\ raft[i][RegionB].logs[next_index].type = LogRollback
/\ raft' = [raft EXCEPT ![i][RegionB].apply_index = next_index]
/\ region' = [region EXCEPT ![i][RegionB] = RegionNormal]
/\ UNCHANGED <<messages, client_vars>>
\* Apply LogNormal.
\* This log simply increases apply_index.
ApplyNormalLog(i, r) ==
LET
next_index == raft[i][r].apply_index + 1
IN
/\ LogAppliable(i, r)
/\ raft[i][r].logs[next_index].type = LogNormal
/\ LET
\* Apply this log if this region is in normal state, otherwise skip it.
\* Notice we don't check for epoch here as what is done in the real
\* world implementation, but these two approaches are equivalent to
\* check whether we have applied PreMergeLog, as applying PreMergeLog
\* will also convert the region state from normal state.
num_applied_delta == IF region[i][r] = RegionNormal THEN 1 ELSE 0
IN
raft' = [raft EXCEPT ![i][r].apply_index = next_index,
![i][r].num_applied = @ + num_applied_delta]
/\ UNCHANGED <<messages, region, client_vars>>
\* Apply Raft logs to make apply_index catch up with commit_index.
ApplyLog(i) ==
\/ \E r \in Region : ApplyNormalLog(i, r)
\/ ApplyPreMergeLog(i)
\/ ApplyMergeLog(i)
\/ ApplyRollbackLog(i)
-------------------------------------------------------------------------------
\* Specification of Raft merge.
Next ==
\* Raft actions.
\/ \E i, j \in Store : \E r \in Region : AppendEntries(i, j, r)
\/ \E i \in Store : \E r \in Region : AdvanceCommitIndex(i, r)
\/ \E m \in messages : Receive(m)
\* External client can send requests to region B leader, to add a new log
\* entry in region B.
\/ \E i \in Store : ClientRequest(i, RegionB, [type |-> LogNormal])
\* Raft merge actions.
\/ ProposeMergeRequest(LeaderB)
\/ PerformRollbackRequest(LeaderB)
\/ \E i \in Store : ApplyLog(i)
Init ==
/\ messages = {}
/\ LET
\* Mark region r on store i as leader.
MarkLeader(rafts, i, r) == [rafts EXCEPT ![i] = [@ EXCEPT ![r].is_leader = TRUE]]
no_leader_raft == [i \in Store |->
[r \in Region |->
[is_leader |-> FALSE,
logs |-> << >>,
commit_index |-> 0,
apply_index |-> 0,
num_applied |-> 0,
match_index |-> [j \in Store |-> 0]
]
]
]
IN
raft = MarkLeader(MarkLeader(no_leader_raft, LeaderA, RegionA),
LeaderB,
RegionB)
/\ region = [i \in Store |-> [r \in Region |-> RegionNormal]]
/\ client_requests_index = 0
Spec ==
Init /\ [][Next]_vars
-------------------------------------------------------------------------------
\* Type invariants.
LogType ==
LET
FlatLogType ==
[type : {LogNormal}]
\cup [type : {LogPreMerge}, min_index : Nat]
\cup [type : {LogRollback}]
IN
FlatLogType
\cup [type : {LogMerge},
min_index : Nat,
commit_index : Nat,
entries : Seq(FlatLogType)]
RaftType ==
[ is_leader : BOOLEAN,
logs : Seq(LogType),
commit_index : Nat,
apply_index : Nat,
num_applied : Nat,
match_index : [Store -> Nat] \* Only available on leader.
\* Initialized to zeroes on followers.
]
RegionType ==
{ RegionNormal, RegionTombStone, RegionMerging }
TypeInvariant ==
/\ raft \in [Store -> [Region -> RaftType]]
/\ region \in [Store -> [Region -> RegionType]]
-------------------------------------------------------------------------------
\* Some invariants for our simplified Raft model.
\* There should be only one leader for each raft group.
OneLeaderInvariant ==
\A r \in Region : Cardinality({s \in Store : raft[s][r].is_leader}) = 1
\* AppendEntries RPC is always sent from leader to follower.
AppendEntriesMessageInvariant ==
\A m \in messages :
LET
r == m.region
leader == CHOOSE s \in Store : raft[s][r].is_leader
IN
/\ m.type = AppendEntriesRequest =>
/\ m.source = leader
\* There should be no gap between the log to append in RPC, and the
\* existing logs.
/\ m.entry_index <= Len(raft[m.dest][r].logs) + 1
/\ m.type = AppendEntriesReply =>
m.dest = leader
\* All known (by leader) replicated logs (logs before match_index) should be on
\* followers, and all replicated logs on followers should be a prefix of logs
\* on leader.
LogInvariant ==
\A r \in Region :
LET
leader == CHOOSE s \in Store : raft[s][r].is_leader
IN
\A s \in Store :
/\ raft[leader][r].match_index[s] <= Len(raft[s][r].logs)
/\ raft[s][r].logs = SubSeq(raft[leader][r].logs, 1, Len(raft[s][r].logs))
\* apply_index should be equal or less than commit_index.
ApplyIndexInvariant ==
\A s \in Store :
\A r \in Region :
raft[s][r].apply_index <= raft[s][r].commit_index
\* Combination of the above invariants.
SimpliedRaftInvariant ==
/\ OneLeaderInvariant
/\ AppendEntriesMessageInvariant
/\ LogInvariant
/\ ApplyIndexInvariant
-------------------------------------------------------------------------------
\* Some invariants for Raft region merge.
\* If a region on two different stores have applied the same logs, they should
\* also share the same region state and number of applied normal logs.
RegionApplyInvariant ==
\A i, j \in Store :
(
/\ i /= j
/\ (\A r \in Region : raft[i][r].apply_index = raft[j][r].apply_index)
) =>
\A r \in Region :
/\ region[i][r] = region[j][r]
/\ raft[i][RegionB].num_applied = raft[j][RegionB].num_applied
\* For any two stores of region B, if both done, they should have the same
\* number of applied logs.
MergeLogInvariant ==
\A i, j \in Store :
(
/\ i /= j
/\ region[i][RegionB] = RegionTombStone
/\ region[j][RegionB] = RegionTombStone
) =>
raft[i][RegionB].num_applied = raft[j][RegionB].num_applied
\* Combination of the above invariants.
RaftMergeInvariant ==
/\ RegionApplyInvariant
/\ MergeLogInvariant
===============================================================================