foundationdb/fdbserver/BlobManager.actor.cpp

1177 lines
40 KiB
C++
Raw Normal View History

/*
* BlobManager.actor.cpp
*
* This source file is part of the FoundationDB open source project
*
* Copyright 2013-2018 Apple Inc. and the FoundationDB project authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
2021-09-03 00:09:37 +08:00
#include <vector>
#include <unordered_map>
#include "fdbclient/BlobWorkerInterface.h"
#include "fdbclient/KeyRangeMap.h"
#include "fdbclient/ReadYourWrites.h"
#include "fdbclient/SystemData.h"
#include "fdbclient/Tuple.h"
#include "fdbserver/BlobManagerInterface.h"
#include "fdbserver/BlobWorker.actor.h"
#include "fdbserver/Knobs.h"
#include "fdbserver/WaitFailure.h"
#include "flow/IRandom.h"
#include "flow/UnitTest.h"
#include "flow/actorcompiler.h" // has to be last include
2021-09-14 23:19:15 +08:00
#include "flow/flow.h"
2021-09-14 23:19:15 +08:00
#define BM_DEBUG false
// TODO add comments + documentation
void handleClientBlobRange(KeyRangeMap<bool>* knownBlobRanges,
Arena ar,
VectorRef<KeyRangeRef>* rangesToAdd,
VectorRef<KeyRangeRef>* rangesToRemove,
KeyRef rangeStart,
KeyRef rangeEnd,
bool rangeActive) {
if (BM_DEBUG) {
printf("db range [%s - %s): %s\n",
rangeStart.printable().c_str(),
rangeEnd.printable().c_str(),
rangeActive ? "T" : "F");
}
KeyRange keyRange(KeyRangeRef(rangeStart, rangeEnd));
auto allRanges = knownBlobRanges->intersectingRanges(keyRange);
for (auto& r : allRanges) {
if (r.value() != rangeActive) {
KeyRef overlapStart = (r.begin() > keyRange.begin) ? r.begin() : keyRange.begin;
KeyRef overlapEnd = (keyRange.end < r.end()) ? keyRange.end : r.end();
KeyRangeRef overlap(overlapStart, overlapEnd);
if (rangeActive) {
if (BM_DEBUG) {
printf("BM Adding client range [%s - %s)\n",
overlapStart.printable().c_str(),
overlapEnd.printable().c_str());
}
rangesToAdd->push_back_deep(ar, overlap);
} else {
if (BM_DEBUG) {
printf("BM Removing client range [%s - %s)\n",
overlapStart.printable().c_str(),
overlapEnd.printable().c_str());
}
rangesToRemove->push_back_deep(ar, overlap);
}
}
}
knownBlobRanges->insert(keyRange, rangeActive);
}
void updateClientBlobRanges(KeyRangeMap<bool>* knownBlobRanges,
RangeResult dbBlobRanges,
Arena ar,
VectorRef<KeyRangeRef>* rangesToAdd,
VectorRef<KeyRangeRef>* rangesToRemove) {
if (BM_DEBUG) {
printf("Updating %d client blob ranges", dbBlobRanges.size() / 2);
for (int i = 0; i < dbBlobRanges.size() - 1; i += 2) {
printf(" [%s - %s)", dbBlobRanges[i].key.printable().c_str(), dbBlobRanges[i + 1].key.printable().c_str());
}
printf("\n");
}
// essentially do merge diff of current known blob ranges and new ranges, to assign new ranges to
// workers and revoke old ranges from workers
// basically, for any range that is set in results that isn't set in ranges, assign the range to the
// worker. for any range that isn't set in results that is set in ranges, revoke the range from the
// worker. and, update ranges to match results as you go
// FIXME: could change this to O(N) instead of O(NLogN) by doing a sorted merge instead of requesting the
// intersection for each insert, but this operation is pretty infrequent so it's probably not necessary
if (dbBlobRanges.size() == 0) {
// special case. Nothing in the DB, reset knownBlobRanges and revoke all existing ranges from workers
handleClientBlobRange(
knownBlobRanges, ar, rangesToAdd, rangesToRemove, normalKeys.begin, normalKeys.end, false);
} else {
if (dbBlobRanges[0].key > normalKeys.begin) {
handleClientBlobRange(
knownBlobRanges, ar, rangesToAdd, rangesToRemove, normalKeys.begin, dbBlobRanges[0].key, false);
}
for (int i = 0; i < dbBlobRanges.size() - 1; i++) {
if (dbBlobRanges[i].key >= normalKeys.end) {
if (BM_DEBUG) {
printf("Found invalid blob range start %s\n", dbBlobRanges[i].key.printable().c_str());
}
break;
}
bool active = dbBlobRanges[i].value == LiteralStringRef("1");
if (active) {
ASSERT(dbBlobRanges[i + 1].value == StringRef());
if (BM_DEBUG) {
printf("BM sees client range [%s - %s)\n",
dbBlobRanges[i].key.printable().c_str(),
dbBlobRanges[i + 1].key.printable().c_str());
}
}
KeyRef endKey = dbBlobRanges[i + 1].key;
if (endKey > normalKeys.end) {
if (BM_DEBUG) {
printf("Removing system keyspace from blob range [%s - %s)\n",
dbBlobRanges[i].key.printable().c_str(),
endKey.printable().c_str());
}
endKey = normalKeys.end;
}
handleClientBlobRange(
knownBlobRanges, ar, rangesToAdd, rangesToRemove, dbBlobRanges[i].key, endKey, active);
}
if (dbBlobRanges[dbBlobRanges.size() - 1].key < normalKeys.end) {
handleClientBlobRange(knownBlobRanges,
ar,
rangesToAdd,
rangesToRemove,
dbBlobRanges[dbBlobRanges.size() - 1].key,
normalKeys.end,
false);
}
}
knownBlobRanges->coalesce(normalKeys);
}
void getRanges(std::vector<std::pair<KeyRangeRef, bool>>& results, KeyRangeMap<bool>& knownBlobRanges) {
if (BM_DEBUG) {
printf("Getting ranges:\n");
}
auto allRanges = knownBlobRanges.ranges();
for (auto& r : allRanges) {
results.emplace_back(r.range(), r.value());
if (BM_DEBUG) {
printf(
" [%s - %s): %s\n", r.begin().printable().c_str(), r.end().printable().c_str(), r.value() ? "T" : "F");
}
}
}
struct RangeAssignmentData {
bool continueAssignment;
RangeAssignmentData() : continueAssignment(false) {}
2021-09-23 01:46:20 +08:00
RangeAssignmentData(bool continueAssignment) : continueAssignment(continueAssignment) {}
};
struct RangeRevokeData {
bool dispose;
RangeRevokeData() {}
RangeRevokeData(bool dispose) : dispose(dispose) {}
};
2021-08-31 02:07:25 +08:00
struct RangeAssignment {
bool isAssign;
KeyRange keyRange;
Optional<UID> worker;
2021-08-31 02:59:53 +08:00
// I tried doing this with a union and it was just kind of messy
Optional<RangeAssignmentData> assign;
Optional<RangeRevokeData> revoke;
2021-08-31 02:07:25 +08:00
};
// TODO: track worker's reads/writes eventually
struct BlobWorkerStats {
int numGranulesAssigned;
BlobWorkerStats(int numGranulesAssigned = 0) : numGranulesAssigned(numGranulesAssigned) {}
};
struct BlobManagerData {
UID id;
Database db;
2021-09-15 23:35:58 +08:00
PromiseStream<Future<Void>> addActor;
2021-09-03 00:09:37 +08:00
std::unordered_map<UID, BlobWorkerInterface> workersById;
std::unordered_map<UID, BlobWorkerStats> workerStats; // mapping between workerID -> workerStats
std::unordered_map<UID, Future<Void>> workerMonitors;
2021-08-31 02:07:25 +08:00
KeyRangeMap<UID> workerAssignments;
KeyRangeMap<bool> knownBlobRanges;
int64_t epoch = -1;
int64_t seqNo = 1;
2021-08-31 02:07:25 +08:00
Promise<Void> iAmReplaced;
// The order maintained here is important. The order ranges are put into the promise stream is the order they get
// assigned sequence numbers
PromiseStream<RangeAssignment> rangesToAssign;
BlobManagerData(UID id, Database db) : id(id), db(db), knownBlobRanges(false, normalKeys.end) {}
~BlobManagerData() { printf("Destroying blob manager data for %s\n", id.toString().c_str()); }
};
ACTOR Future<Standalone<VectorRef<KeyRef>>> splitRange(Reference<ReadYourWritesTransaction> tr, KeyRange range) {
// TODO is it better to just pass empty metrics to estimated?
// TODO handle errors here by pulling out into its own transaction instead of the main loop's transaction, and
// retrying
if (BM_DEBUG) {
printf("Splitting new range [%s - %s)\n", range.begin.printable().c_str(), range.end.printable().c_str());
}
StorageMetrics estimated = wait(tr->getTransaction().getStorageMetrics(range, CLIENT_KNOBS->TOO_MANY));
if (BM_DEBUG) {
printf("Estimated bytes for [%s - %s): %lld\n",
range.begin.printable().c_str(),
range.end.printable().c_str(),
estimated.bytes);
}
if (estimated.bytes > SERVER_KNOBS->BG_SNAPSHOT_FILE_TARGET_BYTES) {
// printf(" Splitting range\n");
// only split on bytes
StorageMetrics splitMetrics;
splitMetrics.bytes = SERVER_KNOBS->BG_SNAPSHOT_FILE_TARGET_BYTES;
splitMetrics.bytesPerKSecond = splitMetrics.infinity;
splitMetrics.iosPerKSecond = splitMetrics.infinity;
splitMetrics.bytesReadPerKSecond = splitMetrics.infinity; // Don't split by readBandwidth
Standalone<VectorRef<KeyRef>> keys =
wait(tr->getTransaction().splitStorageMetrics(range, splitMetrics, estimated));
return keys;
} else {
// printf(" Not splitting range\n");
Standalone<VectorRef<KeyRef>> keys;
keys.push_back_deep(keys.arena(), range.begin);
keys.push_back_deep(keys.arena(), range.end);
return keys;
}
}
// Picks a worker with the fewest number of already assigned ranges.
// If there is a tie, picks one such worker at random.
2021-08-31 02:07:25 +08:00
static UID pickWorkerForAssign(BlobManagerData* bmData) {
int minGranulesAssigned = INT_MAX;
std::vector<UID> eligibleWorkers;
for (auto const& worker : bmData->workerStats) {
UID currId = worker.first;
int granulesAssigned = worker.second.numGranulesAssigned;
if (granulesAssigned < minGranulesAssigned) {
eligibleWorkers.resize(0);
minGranulesAssigned = granulesAssigned;
eligibleWorkers.emplace_back(currId);
} else if (granulesAssigned == minGranulesAssigned) {
eligibleWorkers.emplace_back(currId);
}
}
2021-08-31 02:07:25 +08:00
// pick a random worker out of the eligible workers
ASSERT(eligibleWorkers.size() > 0);
int idx = deterministicRandom()->randomInt(0, eligibleWorkers.size());
if (BM_DEBUG) {
printf("picked worker %s, which has a minimal number (%d) of granules assigned\n",
eligibleWorkers[idx].toString().c_str(),
minGranulesAssigned);
}
return eligibleWorkers[idx];
2021-08-31 02:07:25 +08:00
}
ACTOR Future<Void> doRangeAssignment(BlobManagerData* bmData, RangeAssignment assignment, UID workerID, int64_t seqNo) {
if (BM_DEBUG) {
printf("BM %s %s range [%s - %s) @ (%lld, %lld)\n",
workerID.toString().c_str(),
assignment.isAssign ? "assigning" : "revoking",
assignment.keyRange.begin.printable().c_str(),
assignment.keyRange.end.printable().c_str(),
bmData->epoch,
seqNo);
}
2021-08-31 02:07:25 +08:00
try {
state AssignBlobRangeReply rep;
if (assignment.isAssign) {
ASSERT(assignment.assign.present());
ASSERT(!assignment.revoke.present());
AssignBlobRangeRequest req;
req.keyRange = KeyRangeRef(StringRef(req.arena, assignment.keyRange.begin),
StringRef(req.arena, assignment.keyRange.end));
req.managerEpoch = bmData->epoch;
req.managerSeqno = seqNo;
req.continueAssignment = assignment.assign.get().continueAssignment;
AssignBlobRangeReply _rep = wait(bmData->workersById[workerID].assignBlobRangeRequest.getReply(req));
rep = _rep;
} else {
ASSERT(!assignment.assign.present());
ASSERT(assignment.revoke.present());
RevokeBlobRangeRequest req;
req.keyRange = KeyRangeRef(StringRef(req.arena, assignment.keyRange.begin),
StringRef(req.arena, assignment.keyRange.end));
req.managerEpoch = bmData->epoch;
req.managerSeqno = seqNo;
req.dispose = assignment.revoke.get().dispose;
AssignBlobRangeReply _rep = wait(bmData->workersById[workerID].revokeBlobRangeRequest.getReply(req));
rep = _rep;
}
2021-08-31 02:07:25 +08:00
if (!rep.epochOk) {
if (BM_DEBUG) {
printf("BM heard from BW that there is a new manager with higher epoch\n");
}
2021-08-31 02:07:25 +08:00
if (bmData->iAmReplaced.canBeSet()) {
bmData->iAmReplaced.send(Void());
}
}
} catch (Error& e) {
// TODO confirm: using reliable delivery this should only trigger if the worker is marked as failed, right?
// So assignment needs to be retried elsewhere, and a revoke is trivially complete
if (assignment.isAssign) {
if (BM_DEBUG) {
printf("BM got error assigning range [%s - %s) to worker %s, requeueing\n",
assignment.keyRange.begin.printable().c_str(),
assignment.keyRange.end.printable().c_str());
}
2021-08-31 02:07:25 +08:00
// re-send revoke to queue to handle range being un-assigned from that worker before the new one
RangeAssignment revokeOld;
revokeOld.isAssign = false;
revokeOld.worker = workerID;
revokeOld.keyRange = assignment.keyRange;
revokeOld.revoke = RangeRevokeData(false);
bmData->rangesToAssign.send(revokeOld);
// send assignment back to queue as is, clearing designated worker if present
assignment.worker.reset();
2021-08-31 02:07:25 +08:00
bmData->rangesToAssign.send(assignment);
// FIXME: improvement would be to add history of failed workers to assignment so it can try other ones first
} else {
if (BM_DEBUG) {
printf("BM got error revoking range [%s - %s) from worker %s",
assignment.keyRange.begin.printable().c_str(),
assignment.keyRange.end.printable().c_str());
}
if (assignment.revoke.get().dispose) {
if (BM_DEBUG) {
printf(", retrying for dispose\n");
}
// send assignment back to queue as is, clearing designated worker if present
assignment.worker.reset();
bmData->rangesToAssign.send(assignment);
//
} else {
if (BM_DEBUG) {
printf(", ignoring\n");
}
}
2021-08-31 02:07:25 +08:00
}
}
return Void();
}
2021-08-31 02:07:25 +08:00
ACTOR Future<Void> rangeAssigner(BlobManagerData* bmData) {
loop {
2021-09-14 23:19:15 +08:00
// inject delay into range assignments
if (BUGGIFY_WITH_PROB(0.05)) {
wait(delay(deterministicRandom()->random01()));
}
2021-08-31 02:07:25 +08:00
RangeAssignment assignment = waitNext(bmData->rangesToAssign.getFuture());
int64_t seqNo = bmData->seqNo;
bmData->seqNo++;
// modify the in-memory assignment data structures, and send request off to worker
UID workerId;
if (assignment.isAssign) {
// Ensure range isn't currently assigned anywhere, and there is only 1 intersecting range
auto currentAssignments = bmData->workerAssignments.intersectingRanges(assignment.keyRange);
int count = 0;
for (auto& it : currentAssignments) {
if (assignment.assign.get().continueAssignment) {
ASSERT(assignment.worker.present());
ASSERT(it.value() == assignment.worker.get());
} else {
ASSERT(it.value() == UID());
}
2021-08-31 02:07:25 +08:00
count++;
}
ASSERT(count == 1);
workerId = assignment.worker.present() ? assignment.worker.get() : pickWorkerForAssign(bmData);
2021-08-31 02:07:25 +08:00
bmData->workerAssignments.insert(assignment.keyRange, workerId);
bmData->workerStats[workerId].numGranulesAssigned += 1;
2021-08-31 02:07:25 +08:00
// FIXME: if range is assign, have some sort of semaphore for outstanding assignments so we don't assign
// a ton ranges at once and blow up FDB with reading initial snapshots.
bmData->addActor.send(doRangeAssignment(bmData, assignment, workerId, seqNo));
2021-08-31 02:07:25 +08:00
} else {
// Revoking a range could be a large range that contains multiple ranges.
auto currentAssignments = bmData->workerAssignments.intersectingRanges(assignment.keyRange);
for (auto& it : currentAssignments) {
// ensure range doesn't truncate existing ranges
ASSERT(it.begin() >= assignment.keyRange.begin);
2021-08-31 02:59:53 +08:00
ASSERT(it.end() <= assignment.keyRange.end);
2021-08-31 02:07:25 +08:00
// It is fine for multiple disjoint sub-ranges to have the same sequence number since they were part of
// the same logical change
2021-09-03 00:09:37 +08:00
bmData->workerStats[it.value()].numGranulesAssigned -= 1;
if (!assignment.worker.present() || assignment.worker.get() == it.value())
bmData->addActor.send(doRangeAssignment(bmData, assignment, it.value(), seqNo));
2021-08-31 02:07:25 +08:00
}
bmData->workerAssignments.insert(assignment.keyRange, UID());
}
}
}
ACTOR Future<Void> checkManagerLock(Reference<ReadYourWritesTransaction> tr, BlobManagerData* bmData) {
Optional<Value> currentLockValue = wait(tr->get(blobManagerEpochKey));
ASSERT(currentLockValue.present());
int64_t currentEpoch = decodeBlobManagerEpochValue(currentLockValue.get());
if (currentEpoch != bmData->epoch) {
ASSERT(currentEpoch > bmData->epoch);
2021-09-14 23:19:15 +08:00
if (BM_DEBUG) {
printf("BM %s found new epoch %d > %d in lock check\n",
bmData->id.toString().c_str(),
currentEpoch,
bmData->epoch);
}
if (bmData->iAmReplaced.canBeSet()) {
bmData->iAmReplaced.send(Void());
}
// TODO different error?
throw granule_assignment_conflict();
}
tr->addReadConflictRange(singleKeyRange(blobManagerEpochKey));
return Void();
}
// FIXME: this does all logic in one transaction. Adding a giant range to an existing database to hybridize would spread
// require doing a ton of storage metrics calls, which we should split up across multiple transactions likely.
2021-08-31 02:07:25 +08:00
ACTOR Future<Void> monitorClientRanges(BlobManagerData* bmData) {
loop {
2021-08-31 02:07:25 +08:00
state Reference<ReadYourWritesTransaction> tr = makeReference<ReadYourWritesTransaction>(bmData->db);
if (BM_DEBUG) {
printf("Blob manager checking for range updates\n");
}
loop {
try {
tr->setOption(FDBTransactionOptions::ACCESS_SYSTEM_KEYS);
tr->setOption(FDBTransactionOptions::PRIORITY_SYSTEM_IMMEDIATE);
// TODO probably knobs here? This should always be pretty small though
RangeResult results = wait(krmGetRanges(
tr, blobRangeKeys.begin, KeyRange(normalKeys), 10000, GetRangeLimits::BYTE_LIMIT_UNLIMITED));
ASSERT(!results.more && results.size() < CLIENT_KNOBS->TOO_MANY);
state Arena ar;
ar.dependsOn(results.arena());
VectorRef<KeyRangeRef> rangesToAdd;
VectorRef<KeyRangeRef> rangesToRemove;
// TODO hack for simulation
2021-08-31 02:07:25 +08:00
updateClientBlobRanges(&bmData->knownBlobRanges, results, ar, &rangesToAdd, &rangesToRemove);
for (KeyRangeRef range : rangesToRemove) {
if (BM_DEBUG) {
printf("BM Got range to revoke [%s - %s)\n",
range.begin.printable().c_str(),
range.end.printable().c_str());
}
RangeAssignment ra;
ra.isAssign = false;
ra.keyRange = range;
ra.revoke = RangeRevokeData(true); // dispose=true
bmData->rangesToAssign.send(ra);
}
state std::vector<Future<Standalone<VectorRef<KeyRef>>>> splitFutures;
// Divide new ranges up into equal chunks by using SS byte sample
for (KeyRangeRef range : rangesToAdd) {
2021-08-31 02:07:25 +08:00
// assert that this range contains no currently assigned ranges in this
splitFutures.push_back(splitRange(tr, range));
}
for (auto f : splitFutures) {
Standalone<VectorRef<KeyRef>> splits = wait(f);
if (BM_DEBUG) {
printf("Split client range [%s - %s) into %d ranges:\n",
splits[0].printable().c_str(),
splits[splits.size() - 1].printable().c_str(),
splits.size() - 1);
}
2021-08-31 02:07:25 +08:00
for (int i = 0; i < splits.size() - 1; i++) {
KeyRange range = KeyRange(KeyRangeRef(splits[i], splits[i + 1]));
if (BM_DEBUG) {
printf(" [%s - %s)\n", range.begin.printable().c_str(), range.end.printable().c_str());
}
2021-08-31 02:07:25 +08:00
RangeAssignment ra;
ra.isAssign = true;
ra.keyRange = range;
2021-09-23 01:46:20 +08:00
ra.assign = RangeAssignmentData(false); // continue=false
bmData->rangesToAssign.send(ra);
}
}
state Future<Void> watchFuture = tr->watch(blobRangeChangeKey);
wait(tr->commit());
if (BM_DEBUG) {
printf("Blob manager done processing client ranges, awaiting update\n");
}
2021-08-31 02:07:25 +08:00
wait(watchFuture);
break;
} catch (Error& e) {
if (BM_DEBUG) {
printf("Blob manager got error looking for range updates %s\n", e.name());
}
wait(tr->onError(e));
}
}
}
}
static Key granuleLockKey(KeyRange granuleRange) {
Tuple k;
k.append(granuleRange.begin).append(granuleRange.end);
return k.getDataAsStandalone().withPrefix(blobGranuleLockKeys.begin);
}
// FIXME: propagate errors here
ACTOR Future<Void> maybeSplitRange(BlobManagerData* bmData, UID currentWorkerId, KeyRange range) {
state Reference<ReadYourWritesTransaction> tr = makeReference<ReadYourWritesTransaction>(bmData->db);
state Standalone<VectorRef<KeyRef>> newRanges;
state int64_t newLockSeqno = -1;
// first get ranges to split
loop {
try {
// redo split if previous txn try failed to calculate it
if (newRanges.empty()) {
Standalone<VectorRef<KeyRef>> _newRanges = wait(splitRange(tr, range));
newRanges = _newRanges;
}
break;
} catch (Error& e) {
wait(tr->onError(e));
}
}
if (newRanges.size() == 2) {
// not large enough to split, just reassign back to worker
if (BM_DEBUG) {
printf("Not splitting existing range [%s - %s). Continuing assignment to %s\n",
range.begin.printable().c_str(),
range.end.printable().c_str(),
currentWorkerId.toString().c_str());
}
RangeAssignment raContinue;
raContinue.isAssign = true;
raContinue.worker = currentWorkerId;
raContinue.keyRange = range;
2021-09-23 01:46:20 +08:00
raContinue.assign = RangeAssignmentData(true); // continue assignment and re-snapshot
bmData->rangesToAssign.send(raContinue);
return Void();
}
// Need to split range. Persist intent to split and split metadata to DB BEFORE sending split requests
loop {
try {
tr->reset();
tr->setOption(FDBTransactionOptions::Option::PRIORITY_SYSTEM_IMMEDIATE);
tr->setOption(FDBTransactionOptions::Option::ACCESS_SYSTEM_KEYS);
ASSERT(newRanges.size() >= 2);
// make sure we're still manager when this transaction gets committed
wait(checkManagerLock(tr, bmData));
// acquire lock for old granule to make sure nobody else modifies it
state Key lockKey = granuleLockKey(range);
Optional<Value> lockValue = wait(tr->get(lockKey));
ASSERT(lockValue.present());
std::tuple<int64_t, int64_t, UID> prevGranuleLock = decodeBlobGranuleLockValue(lockValue.get());
if (std::get<0>(prevGranuleLock) > bmData->epoch) {
2021-09-14 23:19:15 +08:00
if (BM_DEBUG) {
printf("BM %s found a higher epoch %d than %d for granule lock of [%s - %s)\n",
bmData->id.toString().c_str(),
std::get<0>(prevGranuleLock),
bmData->epoch,
range.begin.printable().c_str(),
range.end.printable().c_str());
}
if (bmData->iAmReplaced.canBeSet()) {
bmData->iAmReplaced.send(Void());
}
return Void();
}
if (newLockSeqno == -1) {
newLockSeqno = bmData->seqNo;
bmData->seqNo++;
ASSERT(newLockSeqno > std::get<1>(prevGranuleLock));
} else {
// previous transaction could have succeeded but got commit_unknown_result
ASSERT(newLockSeqno >= std::get<1>(prevGranuleLock));
}
2021-09-23 01:46:20 +08:00
// acquire granule lock so nobody else can make changes to this granule.
tr->set(lockKey, blobGranuleLockValueFor(bmData->epoch, newLockSeqno, std::get<2>(prevGranuleLock)));
2021-09-23 01:46:20 +08:00
Standalone<VectorRef<KeyRangeRef>> history;
history.push_back(history.arena(), range);
Value historyValue = blobGranuleHistoryValueFor(history);
// set up split metadata
for (int i = 0; i < newRanges.size() - 1; i++) {
2021-09-23 01:46:20 +08:00
Tuple splitKey;
splitKey.append(range.begin).append(range.end).append(newRanges[i]);
tr->atomicOp(splitKey.getDataAsStandalone().withPrefix(blobGranuleSplitKeys.begin),
blobGranuleSplitValueFor(BlobGranuleSplitState::Started),
MutationRef::SetVersionstampedValue);
Tuple historyKey;
historyKey.append(newRanges[i]).append(newRanges[i + 1]);
tr->set(historyKey.getDataAsStandalone().withPrefix(blobGranuleHistoryKeys.begin), historyValue);
}
2021-09-23 01:46:20 +08:00
wait(tr->commit());
break;
} catch (Error& e) {
if (e.code() == error_code_granule_assignment_conflict) {
if (bmData->iAmReplaced.canBeSet()) {
bmData->iAmReplaced.send(Void());
}
return Void();
}
wait(tr->onError(e));
}
}
if (BM_DEBUG) {
printf("Splitting range [%s - %s) into:\n", range.begin.printable().c_str(), range.end.printable().c_str());
for (int i = 0; i < newRanges.size() - 1; i++) {
printf(" [%s - %s)\n", newRanges[i].printable().c_str(), newRanges[i + 1].printable().c_str());
}
}
// transaction committed, send range assignments
// revoke from current worker
RangeAssignment raRevoke;
raRevoke.isAssign = false;
raRevoke.worker = currentWorkerId;
raRevoke.keyRange = range;
raRevoke.revoke = RangeRevokeData(false); // not a dispose
bmData->rangesToAssign.send(raRevoke);
for (int i = 0; i < newRanges.size() - 1; i++) {
// reassign new range and do handover of previous range
RangeAssignment raAssignSplit;
raAssignSplit.isAssign = true;
raAssignSplit.keyRange = KeyRangeRef(newRanges[i], newRanges[i + 1]);
2021-09-23 01:46:20 +08:00
raAssignSplit.assign = RangeAssignmentData(false);
// don't care who this range gets assigned to
bmData->rangesToAssign.send(raAssignSplit);
}
return Void();
}
ACTOR Future<Void> monitorBlobWorker(BlobManagerData* bmData, BlobWorkerInterface bwInterf) {
try {
state Future<Void> waitFailure = waitFailureClient(bwInterf.waitFailure, SERVER_KNOBS->BLOB_WORKER_TIMEOUT);
state ReplyPromiseStream<GranuleStatusReply> statusStream =
bwInterf.granuleStatusStreamRequest.getReplyStream(GranuleStatusStreamRequest(bmData->epoch));
state KeyRangeMap<std::pair<int64_t, int64_t>> lastSeenSeqno;
loop choose {
when(wait(waitFailure)) {
// FIXME: actually handle this!!
if (BM_DEBUG) {
printf("BM %lld detected BW %s is dead\n", bmData->epoch, bwInterf.id().toString().c_str());
}
return Void();
}
when(GranuleStatusReply _rep = waitNext(statusStream.getFuture())) {
GranuleStatusReply rep = _rep;
if (BM_DEBUG) {
printf("BM %lld got status of [%s - %s) @ (%lld, %lld) from BW %s: %s\n",
bmData->epoch,
rep.granuleRange.begin.printable().c_str(),
rep.granuleRange.end.printable().c_str(),
rep.epoch,
rep.seqno,
bwInterf.id().toString().c_str(),
rep.doSplit ? "split" : "");
}
if (rep.epoch > bmData->epoch) {
if (BM_DEBUG) {
printf("BM heard from BW that there is a new manager with higher epoch\n");
}
if (bmData->iAmReplaced.canBeSet()) {
bmData->iAmReplaced.send(Void());
}
}
// TODO maybe this won't be true eventually, but right now the only time the blob worker reports back is
// to split the range.
ASSERT(rep.doSplit);
// FIXME: only evaluate for split if this worker currently owns the granule in this blob manager's
// mapping
auto lastReqForGranule = lastSeenSeqno.rangeContaining(rep.granuleRange.begin);
if (rep.granuleRange.begin == lastReqForGranule.begin() &&
rep.granuleRange.end == lastReqForGranule.end() && rep.epoch == lastReqForGranule.value().first &&
rep.seqno == lastReqForGranule.value().second) {
if (BM_DEBUG) {
printf("Manager %lld received repeat status for the same granule [%s - %s) @ %lld, ignoring.",
bmData->epoch,
rep.granuleRange.begin.printable().c_str(),
rep.granuleRange.end.printable().c_str());
}
} else {
if (BM_DEBUG) {
printf("Manager %lld evaluating [%s - %s) for split\n",
bmData->epoch,
rep.granuleRange.begin.printable().c_str(),
rep.granuleRange.end.printable().c_str());
}
lastSeenSeqno.insert(rep.granuleRange, std::pair(rep.epoch, rep.seqno));
bmData->addActor.send(maybeSplitRange(bmData, bwInterf.id(), rep.granuleRange));
}
}
}
} catch (Error& e) {
// FIXME: forward errors somewhere from here
if (BM_DEBUG) {
printf("BM got unexpected error %s monitoring BW %s\n", e.name(), bwInterf.id().toString().c_str());
}
throw e;
}
}
2021-08-31 02:59:53 +08:00
// TODO this is only for chaos testing right now!! REMOVE LATER
ACTOR Future<Void> rangeMover(BlobManagerData* bmData) {
loop {
wait(delay(30.0));
if (bmData->workersById.size() > 1) {
int tries = 10;
while (tries > 0) {
tries--;
auto randomRange = bmData->workerAssignments.randomRange();
if (randomRange.value() != UID()) {
if (BM_DEBUG) {
printf("Range mover moving range [%s - %s): %s\n",
randomRange.begin().printable().c_str(),
randomRange.end().printable().c_str(),
randomRange.value().toString().c_str());
}
2021-08-31 02:59:53 +08:00
RangeAssignment revokeOld;
revokeOld.isAssign = false;
revokeOld.keyRange = randomRange.range();
revokeOld.worker = randomRange.value();
revokeOld.revoke = RangeRevokeData(false);
bmData->rangesToAssign.send(revokeOld);
RangeAssignment assignNew;
assignNew.isAssign = true;
assignNew.keyRange = randomRange.range();
2021-09-23 01:46:20 +08:00
assignNew.assign = RangeAssignmentData(false); // not a continue
bmData->rangesToAssign.send(assignNew);
2021-08-31 02:59:53 +08:00
break;
}
}
if (tries == 0 && BM_DEBUG) {
2021-09-23 01:46:20 +08:00
printf("Range mover couldn't find random range to move, skipping\n");
2021-08-31 02:59:53 +08:00
}
} else if (BM_DEBUG) {
2021-08-31 02:59:53 +08:00
printf("Range mover found %d workers, skipping\n", bmData->workerAssignments.size());
}
}
}
2021-09-15 23:35:58 +08:00
ACTOR Future<Void> blobManager(BlobManagerInterface bmInterf,
Reference<AsyncVar<ServerDBInfo> const> dbInfo,
int64_t epoch) {
2021-08-31 02:07:25 +08:00
state BlobManagerData self(deterministicRandom()->randomUniqueID(),
openDBOnServer(dbInfo, TaskPriority::DefaultEndpoint, LockAware::True));
state Future<Void> collection = actorCollection(self.addActor.getFuture());
2021-08-31 02:07:25 +08:00
if (BM_DEBUG) {
printf("Blob manager starting...\n");
}
2021-09-15 23:35:58 +08:00
self.epoch = epoch;
// make sure the epoch hasn't gotten stale
state Reference<ReadYourWritesTransaction> tr = makeReference<ReadYourWritesTransaction>(self.db);
tr->setOption(FDBTransactionOptions::ACCESS_SYSTEM_KEYS);
tr->setOption(FDBTransactionOptions::PRIORITY_SYSTEM_IMMEDIATE);
try {
wait(checkManagerLock(tr, &self));
} catch (Error& e) {
if (BM_DEBUG) {
printf("Blob manager lock check got unexpected error %s. Dying...\n", e.name());
2021-09-15 23:35:58 +08:00
}
return Void();
2021-09-15 23:35:58 +08:00
}
if (BM_DEBUG) {
2021-09-15 23:35:58 +08:00
printf("Blob manager acquired lock at epoch %lld\n", epoch);
}
2021-08-31 02:07:25 +08:00
int numWorkers = 2;
for (int i = 0; i < numWorkers; i++) {
2021-09-15 23:35:58 +08:00
state BlobWorkerInterface bwInterf(bmInterf.locality, deterministicRandom()->randomUniqueID());
2021-08-31 02:07:25 +08:00
bwInterf.initEndpoints();
self.workersById.insert({ bwInterf.id(), bwInterf });
self.workerStats.insert({ bwInterf.id(), BlobWorkerStats() });
self.workerMonitors.insert({ bwInterf.id(), monitorBlobWorker(&self, bwInterf) });
self.addActor.send(blobWorker(bwInterf, dbInfo));
2021-08-31 02:07:25 +08:00
}
self.addActor.send(monitorClientRanges(&self));
self.addActor.send(rangeAssigner(&self));
2021-09-23 01:46:20 +08:00
if (BUGGIFY) {
self.addActor.send(rangeMover(&self));
2021-09-23 01:46:20 +08:00
}
2021-08-31 02:07:25 +08:00
// TODO probably other things here eventually
try {
loop choose {
when(wait(self.iAmReplaced.getFuture())) {
if (BM_DEBUG) {
printf("Blob Manager exiting because it is replaced\n");
}
break;
}
when(HaltBlobManagerRequest req = waitNext(bmInterf.haltBlobManager.getFuture())) {
req.reply.send(Void());
TraceEvent("BlobManagerHalted", bmInterf.id()).detail("ReqID", req.requesterID);
break;
}
when(wait(collection)) {
TraceEvent("BlobManagerActorCollectionError");
ASSERT(false);
throw internal_error();
}
2021-08-31 02:07:25 +08:00
}
} catch (Error& err) {
TraceEvent("BlobManagerDied", bmInterf.id()).error(err, true);
2021-08-31 02:07:25 +08:00
}
return Void();
2021-08-31 02:07:25 +08:00
}
// Test:
// start empty
// DB has [A - B). That should show up in knownBlobRanges and should be in added
// DB has nothing. knownBlobRanges should be empty and [A - B) should be in removed
// DB has [A - B) and [C - D). They should both show up in knownBlobRanges and added.
// DB has [A - D). It should show up coalesced in knownBlobRanges, and [B - C) should be in added.
// DB has [A - C). It should show up coalesced in knownBlobRanges, and [C - D) should be in removed.
// DB has [B - C). It should show up coalesced in knownBlobRanges, and [A - B) should be removed.
// DB has [B - D). It should show up coalesced in knownBlobRanges, and [C - D) should be removed.
// DB has [A - D). It should show up coalesced in knownBlobRanges, and [A - B) should be removed.
// DB has [A - B) and [C - D). They should show up in knownBlobRanges, and [B - C) should be in removed.
2021-09-15 23:35:58 +08:00
// DB has [B - C). It should show up in knownBlobRanges, [B - C) should be in added, and [A - B) and [C - D) should
// be in removed.
TEST_CASE("/blobmanager/updateranges") {
KeyRangeMap<bool> knownBlobRanges(false, normalKeys.end);
Arena ar;
VectorRef<KeyRangeRef> added;
VectorRef<KeyRangeRef> removed;
StringRef active = LiteralStringRef("1");
StringRef inactive = StringRef();
RangeResult dbDataEmpty;
vector<std::pair<KeyRangeRef, bool>> kbrRanges;
StringRef keyA = StringRef(ar, LiteralStringRef("A"));
StringRef keyB = StringRef(ar, LiteralStringRef("B"));
StringRef keyC = StringRef(ar, LiteralStringRef("C"));
StringRef keyD = StringRef(ar, LiteralStringRef("D"));
// db data setup
RangeResult dbDataAB;
dbDataAB.emplace_back(ar, keyA, active);
dbDataAB.emplace_back(ar, keyB, inactive);
RangeResult dbDataAC;
dbDataAC.emplace_back(ar, keyA, active);
dbDataAC.emplace_back(ar, keyC, inactive);
RangeResult dbDataAD;
dbDataAD.emplace_back(ar, keyA, active);
dbDataAD.emplace_back(ar, keyD, inactive);
RangeResult dbDataBC;
dbDataBC.emplace_back(ar, keyB, active);
dbDataBC.emplace_back(ar, keyC, inactive);
RangeResult dbDataBD;
dbDataBD.emplace_back(ar, keyB, active);
dbDataBD.emplace_back(ar, keyD, inactive);
RangeResult dbDataCD;
dbDataCD.emplace_back(ar, keyC, active);
dbDataCD.emplace_back(ar, keyD, inactive);
RangeResult dbDataAB_CD;
dbDataAB_CD.emplace_back(ar, keyA, active);
dbDataAB_CD.emplace_back(ar, keyB, inactive);
dbDataAB_CD.emplace_back(ar, keyC, active);
dbDataAB_CD.emplace_back(ar, keyD, inactive);
// key ranges setup
KeyRangeRef rangeAB = KeyRangeRef(keyA, keyB);
KeyRangeRef rangeAC = KeyRangeRef(keyA, keyC);
KeyRangeRef rangeAD = KeyRangeRef(keyA, keyD);
KeyRangeRef rangeBC = KeyRangeRef(keyB, keyC);
KeyRangeRef rangeBD = KeyRangeRef(keyB, keyD);
KeyRangeRef rangeCD = KeyRangeRef(keyC, keyD);
KeyRangeRef rangeStartToA = KeyRangeRef(normalKeys.begin, keyA);
KeyRangeRef rangeStartToB = KeyRangeRef(normalKeys.begin, keyB);
KeyRangeRef rangeStartToC = KeyRangeRef(normalKeys.begin, keyC);
KeyRangeRef rangeBToEnd = KeyRangeRef(keyB, normalKeys.end);
KeyRangeRef rangeCToEnd = KeyRangeRef(keyC, normalKeys.end);
KeyRangeRef rangeDToEnd = KeyRangeRef(keyD, normalKeys.end);
// actual test
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 1);
ASSERT(kbrRanges[0].first == normalKeys);
ASSERT(!kbrRanges[0].second);
// DB has [A - B)
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataAB, ar, &added, &removed);
ASSERT(added.size() == 1);
ASSERT(added[0] == rangeAB);
ASSERT(removed.size() == 0);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 3);
ASSERT(kbrRanges[0].first == rangeStartToA);
ASSERT(!kbrRanges[0].second);
ASSERT(kbrRanges[1].first == rangeAB);
ASSERT(kbrRanges[1].second);
ASSERT(kbrRanges[2].first == rangeBToEnd);
ASSERT(!kbrRanges[2].second);
// DB has nothing
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataEmpty, ar, &added, &removed);
ASSERT(added.size() == 0);
ASSERT(removed.size() == 1);
ASSERT(removed[0] == rangeAB);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges[0].first == normalKeys);
ASSERT(!kbrRanges[0].second);
// DB has [A - B) and [C - D)
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataAB_CD, ar, &added, &removed);
ASSERT(added.size() == 2);
ASSERT(added[0] == rangeAB);
ASSERT(added[1] == rangeCD);
ASSERT(removed.size() == 0);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 5);
ASSERT(kbrRanges[0].first == rangeStartToA);
ASSERT(!kbrRanges[0].second);
ASSERT(kbrRanges[1].first == rangeAB);
ASSERT(kbrRanges[1].second);
ASSERT(kbrRanges[2].first == rangeBC);
ASSERT(!kbrRanges[2].second);
ASSERT(kbrRanges[3].first == rangeCD);
ASSERT(kbrRanges[3].second);
ASSERT(kbrRanges[4].first == rangeDToEnd);
ASSERT(!kbrRanges[4].second);
// DB has [A - D)
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataAD, ar, &added, &removed);
ASSERT(added.size() == 1);
ASSERT(added[0] == rangeBC);
ASSERT(removed.size() == 0);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 3);
ASSERT(kbrRanges[0].first == rangeStartToA);
ASSERT(!kbrRanges[0].second);
ASSERT(kbrRanges[1].first == rangeAD);
ASSERT(kbrRanges[1].second);
ASSERT(kbrRanges[2].first == rangeDToEnd);
ASSERT(!kbrRanges[2].second);
// DB has [A - C)
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataAC, ar, &added, &removed);
ASSERT(added.size() == 0);
ASSERT(removed.size() == 1);
ASSERT(removed[0] == rangeCD);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 3);
ASSERT(kbrRanges[0].first == rangeStartToA);
ASSERT(!kbrRanges[0].second);
ASSERT(kbrRanges[1].first == rangeAC);
ASSERT(kbrRanges[1].second);
ASSERT(kbrRanges[2].first == rangeCToEnd);
ASSERT(!kbrRanges[2].second);
// DB has [B - C)
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataBC, ar, &added, &removed);
ASSERT(added.size() == 0);
ASSERT(removed.size() == 1);
ASSERT(removed[0] == rangeAB);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 3);
ASSERT(kbrRanges[0].first == rangeStartToB);
ASSERT(!kbrRanges[0].second);
ASSERT(kbrRanges[1].first == rangeBC);
ASSERT(kbrRanges[1].second);
ASSERT(kbrRanges[2].first == rangeCToEnd);
ASSERT(!kbrRanges[2].second);
// DB has [B - D)
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataBD, ar, &added, &removed);
ASSERT(added.size() == 1);
ASSERT(added[0] == rangeCD);
ASSERT(removed.size() == 0);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 3);
ASSERT(kbrRanges[0].first == rangeStartToB);
ASSERT(!kbrRanges[0].second);
ASSERT(kbrRanges[1].first == rangeBD);
ASSERT(kbrRanges[1].second);
ASSERT(kbrRanges[2].first == rangeDToEnd);
ASSERT(!kbrRanges[2].second);
// DB has [A - D)
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataAD, ar, &added, &removed);
ASSERT(added.size() == 1);
ASSERT(added[0] == rangeAB);
ASSERT(removed.size() == 0);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 3);
ASSERT(kbrRanges[0].first == rangeStartToA);
ASSERT(!kbrRanges[0].second);
ASSERT(kbrRanges[1].first == rangeAD);
ASSERT(kbrRanges[1].second);
ASSERT(kbrRanges[2].first == rangeDToEnd);
ASSERT(!kbrRanges[2].second);
// DB has [A - B) and [C - D)
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataAB_CD, ar, &added, &removed);
ASSERT(added.size() == 0);
ASSERT(removed.size() == 1);
ASSERT(removed[0] == rangeBC);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 5);
ASSERT(kbrRanges[0].first == rangeStartToA);
ASSERT(!kbrRanges[0].second);
ASSERT(kbrRanges[1].first == rangeAB);
ASSERT(kbrRanges[1].second);
ASSERT(kbrRanges[2].first == rangeBC);
ASSERT(!kbrRanges[2].second);
ASSERT(kbrRanges[3].first == rangeCD);
ASSERT(kbrRanges[3].second);
ASSERT(kbrRanges[4].first == rangeDToEnd);
ASSERT(!kbrRanges[4].second);
// DB has [B - C)
kbrRanges.clear();
added.clear();
removed.clear();
updateClientBlobRanges(&knownBlobRanges, dbDataBC, ar, &added, &removed);
ASSERT(added.size() == 1);
ASSERT(added[0] == rangeBC);
ASSERT(removed.size() == 2);
ASSERT(removed[0] == rangeAB);
ASSERT(removed[1] == rangeCD);
getRanges(kbrRanges, knownBlobRanges);
ASSERT(kbrRanges.size() == 3);
ASSERT(kbrRanges[0].first == rangeStartToB);
ASSERT(!kbrRanges[0].second);
ASSERT(kbrRanges[1].first == rangeBC);
ASSERT(kbrRanges[1].second);
ASSERT(kbrRanges[2].first == rangeCToEnd);
ASSERT(!kbrRanges[2].second);
return Void();
}