Coverage Report

Created: 2026-08-07 14:29

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
be/src/storage/rowset_version_mgr.cpp
Line
Count
Source
1
// Licensed to the Apache Software Foundation (ASF) under one
2
// or more contributor license agreements.  See the NOTICE file
3
// distributed with this work for additional information
4
// regarding copyright ownership.  The ASF licenses this file
5
// to you under the Apache License, Version 2.0 (the
6
// "License"); you may not use this file except in compliance
7
// with the License.  You may obtain a copy of the License at
8
//
9
//   http://www.apache.org/licenses/LICENSE-2.0
10
//
11
// Unless required by applicable law or agreed to in writing,
12
// software distributed under the License is distributed on an
13
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14
// KIND, either express or implied.  See the License for the
15
// specific language governing permissions and limitations
16
// under the License.
17
18
#include <brpc/controller.h>
19
#include <bthread/bthread.h>
20
#include <bthread/countdown_event.h>
21
#include <bthread/mutex.h>
22
#include <bthread/types.h>
23
#include <bvar/latency_recorder.h>
24
#include <gen_cpp/FrontendService_types.h>
25
#include <gen_cpp/HeartbeatService_types.h>
26
#include <gen_cpp/Types_types.h>
27
#include <gen_cpp/internal_service.pb.h>
28
#include <gen_cpp/olap_file.pb.h>
29
#include <glog/logging.h>
30
31
#include <cstdint>
32
#include <memory>
33
#include <mutex>
34
#include <optional>
35
#include <ranges>
36
#include <sstream>
37
#include <utility>
38
39
#include "cloud/config.h"
40
#include "common/status.h"
41
#include "cpp/sync_point.h"
42
#include "runtime/cluster_info.h"
43
#include "service/backend_options.h"
44
#include "service/internal_service.h"
45
#include "storage/olap_common.h"
46
#include "storage/rowset/rowset.h"
47
#include "storage/rowset/rowset_factory.h"
48
#include "storage/rowset/rowset_reader.h"
49
#include "storage/tablet/base_tablet.h"
50
#include "util/brpc_client_cache.h"
51
#include "util/client_cache.h"
52
#include "util/debug_points.h"
53
#include "util/thrift_rpc_helper.h"
54
#include "util/time.h"
55
56
namespace doris {
57
58
using namespace ErrorCode;
59
using namespace std::ranges;
60
61
static bvar::LatencyRecorder g_remote_fetch_tablet_rowsets_single_request_latency(
62
        "remote_fetch_rowsets_single_rpc");
63
static bvar::LatencyRecorder g_remote_fetch_tablet_rowsets_latency("remote_fetch_rowsets");
64
65
[[nodiscard]] Result<std::vector<Version>> BaseTablet::capture_consistent_versions_unlocked(
66
463k
        const Version& version_range, const CaptureRowsetOps& options) const {
67
463k
    std::vector<Version> version_path;
68
463k
    auto st =
69
463k
            _timestamped_version_tracker.capture_consistent_versions(version_range, &version_path);
70
463k
    if (!st && !options.quiet) {
71
2
        auto missed_versions = get_missed_versions_unlocked(version_range.second);
72
2
        if (missed_versions.empty()) {
73
0
            LOG(WARNING) << fmt::format(
74
0
                    "version already has been merged. version_range={}, max_version={}, "
75
0
                    "tablet_id={}",
76
0
                    version_range.to_string(), _tablet_meta->max_version().second, tablet_id());
77
0
            return ResultError(Status::Error<VERSION_ALREADY_MERGED>(
78
0
                    "missed versions is empty, version_range={}, max_version={}, tablet_id={}",
79
0
                    version_range.to_string(), _tablet_meta->max_version().second, tablet_id()));
80
0
        }
81
2
        LOG(WARNING) << fmt::format("missed version for version_range={}, tablet_id={}, st={}",
82
2
                                    version_range.to_string(), tablet_id(), st);
83
2
        _print_missed_versions(missed_versions);
84
2
        if (!options.skip_missing_versions) {
85
2
            return ResultError(std::move(st));
86
2
        }
87
2
        LOG(WARNING) << "force skipping missing version for tablet:" << tablet_id();
88
0
    }
89
463k
    DBUG_EXECUTE_IF("Tablet::capture_consistent_versions.inject_failure", {
90
463k
        auto tablet_id = dp->param<int64_t>("tablet_id", -1);
91
463k
        auto skip_by_option = dp->param<bool>("skip_by_option", false);
92
463k
        if (skip_by_option && !options.enable_fetch_rowsets_from_peers) {
93
463k
            return version_path;
94
463k
        }
95
463k
        if ((tablet_id != -1 && (tablet_id == _tablet_meta->tablet_id())) || tablet_id == -2) {
96
463k
            return ResultError(Status::Error<VERSION_ALREADY_MERGED>("version already merged"));
97
463k
        }
98
463k
    });
99
463k
    return version_path;
100
463k
}
101
102
[[nodiscard]] Result<CaptureRowsetResult> BaseTablet::capture_consistent_rowsets_unlocked(
103
458k
        const Version& version_range, const CaptureRowsetOps& options) const {
104
458k
    CaptureRowsetResult result;
105
458k
    auto& rowsets = result.rowsets;
106
458k
    auto maybe_versions = capture_consistent_versions_unlocked(version_range, options);
107
459k
    if (maybe_versions) {
108
459k
        const auto& version_paths = maybe_versions.value();
109
459k
        rowsets.reserve(version_paths.size());
110
111
459k
        auto rowset_for_version = [&](const Version& version,
112
1.64M
                                      bool include_stale) -> Result<RowsetSharedPtr> {
113
1.64M
            if (auto it = _rs_version_map.find(version); it != _rs_version_map.end()) {
114
1.64M
                return it->second;
115
18.4E
            } else {
116
18.4E
                VLOG_NOTICE << "fail to find Rowset in rs_version for version. tablet="
117
18.4E
                            << tablet_id() << ", version='" << version.first << "-"
118
18.4E
                            << version.second;
119
18.4E
            }
120
18.4E
            if (include_stale) {
121
58
                if (auto it = _stale_rs_version_map.find(version);
122
58
                    it != _stale_rs_version_map.end()) {
123
58
                    return it->second;
124
58
                } else {
125
0
                    LOG(WARNING) << fmt::format(
126
0
                            "fail to find Rowset in stale_rs_version for version. tablet={}, "
127
0
                            "version={}-{}",
128
0
                            tablet_id(), version.first, version.second);
129
0
                }
130
58
            }
131
18.4E
            return ResultError(Status::Error<CAPTURE_ROWSET_ERROR>(
132
18.4E
                    "failed to find rowset for version={}", version.to_string()));
133
18.4E
        };
134
135
1.64M
        for (const auto& version : version_paths) {
136
1.64M
            auto ret = rowset_for_version(version, options.include_stale_rowsets);
137
1.64M
            if (!ret) {
138
0
                return ResultError(std::move(ret.error()));
139
0
            }
140
141
1.64M
            rowsets.push_back(std::move(ret.value()));
142
1.64M
        }
143
459k
        if (need_read_delete_bitmap()) {
144
417k
            result.delete_bitmap = _tablet_meta->delete_bitmap_ptr();
145
417k
        }
146
459k
        return result;
147
459k
    }
148
149
18.4E
    if (!config::is_cloud_mode() || !options.enable_fetch_rowsets_from_peers) {
150
2
        return ResultError(std::move(maybe_versions.error()));
151
2
    }
152
18.4E
    auto ret = _remote_capture_rowsets(version_range);
153
18.4E
    if (!ret) {
154
0
        auto st = Status::Error<VERSION_ALREADY_MERGED>(
155
0
                "version already merged, meet error during remote capturing rowsets, "
156
0
                "error={}, version_range={}",
157
0
                ret.error().to_string(), version_range.to_string());
158
0
        return ResultError(std::move(st));
159
0
    }
160
18.4E
    return ret;
161
18.4E
}
162
163
[[nodiscard]] Result<std::vector<RowSetSplits>> BaseTablet::capture_rs_readers_unlocked(
164
5.00k
        const Version& version_range, const CaptureRowsetOps& options) const {
165
5.00k
    auto maybe_rs_list = capture_consistent_rowsets_unlocked(version_range, options);
166
5.00k
    if (!maybe_rs_list) {
167
1
        return ResultError(std::move(maybe_rs_list.error()));
168
1
    }
169
4.99k
    const auto& rs_list = maybe_rs_list.value().rowsets;
170
4.99k
    std::vector<RowSetSplits> rs_splits;
171
4.99k
    rs_splits.reserve(rs_list.size());
172
9.35k
    for (const auto& rs : rs_list) {
173
9.35k
        RowsetReaderSharedPtr rs_reader;
174
9.35k
        auto st = rs->create_reader(&rs_reader);
175
9.35k
        if (!st) {
176
0
            return ResultError(Status::Error<CAPTURE_ROWSET_READER_ERROR>(
177
0
                    "failed to create reader for rowset={}, reason={}", rs->rowset_id().to_string(),
178
0
                    st.to_string()));
179
0
        }
180
9.35k
        rs_splits.emplace_back(std::move(rs_reader));
181
9.35k
    }
182
4.99k
    return rs_splits;
183
4.99k
}
184
185
[[nodiscard]] Result<TabletReadSource> BaseTablet::capture_read_source(
186
1.48M
        const Version& version_range, const CaptureRowsetOps& options) {
187
1.48M
    std::shared_lock rdlock(get_header_lock());
188
1.48M
    auto maybe_result = capture_consistent_rowsets_unlocked(version_range, options);
189
1.48M
    if (!maybe_result) {
190
1
        return ResultError(std::move(maybe_result.error()));
191
1
    }
192
1.48M
    auto rowsets_result = std::move(maybe_result.value());
193
1.48M
    TabletReadSource read_source;
194
1.48M
    read_source.delete_bitmap = std::move(rowsets_result.delete_bitmap);
195
1.48M
    const auto& rowsets = rowsets_result.rowsets;
196
1.48M
    read_source.rs_splits.reserve(rowsets.size());
197
6.20M
    for (const auto& rs : rowsets) {
198
6.20M
        RowsetReaderSharedPtr rs_reader;
199
6.20M
        auto st = rs->create_reader(&rs_reader);
200
6.20M
        if (!st) {
201
0
            return ResultError(Status::Error<CAPTURE_ROWSET_READER_ERROR>(
202
0
                    "failed to create reader for rowset={}, reason={}", rs->rowset_id().to_string(),
203
0
                    st.to_string()));
204
0
        }
205
6.20M
        read_source.rs_splits.emplace_back(std::move(rs_reader));
206
6.20M
    }
207
1.48M
    return read_source;
208
1.48M
}
209
210
template <typename Fn, typename... Args>
211
0
bool call_bthread(bthread_t& th, const bthread_attr_t* attr, Fn&& fn, Args&&... args) {
212
0
    auto p_wrap_fn = new auto([=] { fn(args...); });
213
0
    auto call_back = [](void* ar) -> void* {
214
0
        auto f = reinterpret_cast<decltype(p_wrap_fn)>(ar);
215
0
        (*f)();
216
0
        delete f;
217
0
        return nullptr;
218
0
    };
219
0
    return bthread_start_background(&th, attr, call_back, p_wrap_fn) == 0;
220
0
}
221
222
struct GetRowsetsCntl : std::enable_shared_from_this<GetRowsetsCntl> {
223
    struct RemoteGetRowsetResult {
224
        std::vector<RowsetMetaSharedPtr> rowsets;
225
        std::unique_ptr<DeleteBitmap> delete_bitmap;
226
    };
227
228
0
    Status start_req_bg() {
229
0
        task_cnt = req_addrs.size();
230
0
        for (const auto& [ip, port] : req_addrs) {
231
0
            bthread_t tid;
232
0
            bthread_attr_t attr = BTHREAD_ATTR_NORMAL;
233
234
0
            bool succ = call_bthread(tid, &attr, [self = shared_from_this(), &ip, port]() {
235
0
                LOG(INFO) << "start to get tablet rowsets from peer BE, ip=" << ip;
236
0
                Defer defer_log {[&ip, port]() {
237
0
                    LOG(INFO) << "finish to get rowsets from peer BE, ip=" << ip
238
0
                              << ", port=" << port;
239
0
                }};
240
241
0
                PGetTabletRowsetsRequest req;
242
0
                req.set_tablet_id(self->tablet_id);
243
0
                req.set_version_start(self->version_range.first);
244
0
                req.set_version_end(self->version_range.second);
245
0
                if (self->delete_bitmap_keys.has_value()) {
246
0
                    req.mutable_delete_bitmap_keys()->CopyFrom(self->delete_bitmap_keys.value());
247
0
                }
248
0
                brpc::Controller cntl;
249
0
                cntl.set_timeout_ms(60000);
250
0
                cntl.set_max_retry(3);
251
0
                PGetTabletRowsetsResponse response;
252
0
                auto start_tm_us = MonotonicMicros();
253
0
#ifndef BE_TEST
254
0
                std::shared_ptr<PBackendService_Stub> stub =
255
0
                        ExecEnv::GetInstance()->brpc_internal_client_cache()->get_client(ip, port);
256
0
                if (stub == nullptr) {
257
0
                    self->result = ResultError(Status::InternalError(
258
0
                            "failed to fetch get_tablet_rowsets stub, ip={}, port={}", ip, port));
259
0
                    return;
260
0
                }
261
0
                stub->get_tablet_rowsets(&cntl, &req, &response, nullptr);
262
#else
263
                TEST_SYNC_POINT_CALLBACK("get_tablet_rowsets", &response);
264
#endif
265
0
                g_remote_fetch_tablet_rowsets_single_request_latency
266
0
                        << MonotonicMicros() - start_tm_us;
267
268
0
                std::unique_lock l(self->butex);
269
0
                if (self->done) {
270
0
                    return;
271
0
                }
272
0
                --self->task_cnt;
273
0
                auto resp_st = Status::create(response.status());
274
0
                DBUG_EXECUTE_IF("GetRowsetCntl::start_req_bg.inject_failure",
275
0
                                { resp_st = Status::InternalError("inject error"); });
276
0
                if (cntl.Failed() || !resp_st) {
277
0
                    if (self->task_cnt != 0) {
278
0
                        return;
279
0
                    }
280
0
                    std::stringstream reason;
281
0
                    reason << "failed to get rowsets from all replicas, tablet_id="
282
0
                           << self->tablet_id;
283
0
                    if (cntl.Failed()) {
284
0
                        reason << ", reason=[" << cntl.ErrorCode() << "] " << cntl.ErrorText();
285
0
                    } else {
286
0
                        reason << ", reason=" << resp_st.to_string();
287
0
                    }
288
0
                    self->result = ResultError(Status::InternalError(reason.str()));
289
0
                    self->done = true;
290
0
                    self->event.signal();
291
0
                    return;
292
0
                }
293
294
0
                Defer done_cb {[&]() {
295
0
                    self->done = true;
296
0
                    self->event.signal();
297
0
                }};
298
0
                std::vector<RowsetMetaSharedPtr> rs_metas;
299
0
                for (auto&& rs_pb : response.rowsets()) {
300
0
                    auto rs_meta = std::make_shared<RowsetMeta>();
301
0
                    if (!rs_meta->init_from_pb(rs_pb)) {
302
0
                        self->result =
303
0
                                ResultError(Status::InternalError("failed to init rowset from pb"));
304
0
                        return;
305
0
                    }
306
0
                    rs_metas.push_back(std::move(rs_meta));
307
0
                }
308
0
                CaptureRowsetResult result;
309
0
                self->result->rowsets = std::move(rs_metas);
310
311
0
                if (response.has_delete_bitmap()) {
312
0
                    self->result->delete_bitmap = std::make_unique<DeleteBitmap>(
313
0
                            DeleteBitmap::from_pb(response.delete_bitmap(), self->tablet_id));
314
0
                }
315
0
            });
316
317
0
            if (!succ) {
318
0
                return Status::InternalError(
319
0
                        "failed to create bthread when request rowsets for tablet={}", tablet_id);
320
0
            }
321
0
        }
322
0
        return Status::OK();
323
0
    }
324
325
0
    Result<RemoteGetRowsetResult> wait_for_ret() {
326
0
        event.wait();
327
0
        return std::move(result);
328
0
    }
329
330
    int64_t tablet_id;
331
    std::vector<std::pair<std::string, int32_t>> req_addrs;
332
    Version version_range;
333
    std::optional<DeleteBitmapPB> delete_bitmap_keys = std::nullopt;
334
335
private:
336
    size_t task_cnt;
337
338
    bthread::Mutex butex;
339
    bthread::CountdownEvent event {1};
340
    bool done = false;
341
342
    Result<RemoteGetRowsetResult> result;
343
};
344
345
Result<std::vector<std::pair<std::string, int32_t>>> get_peer_replicas_addresses(
346
0
        const int64_t tablet_id) {
347
0
    auto* cluster_info = ExecEnv::GetInstance()->cluster_info();
348
0
    DCHECK_NE(cluster_info, nullptr);
349
0
    auto master_addr = cluster_info->master_fe_addr;
350
0
    TGetTabletReplicaInfosRequest req;
351
0
    req.tablet_ids.push_back(tablet_id);
352
0
    TGetTabletReplicaInfosResult resp;
353
0
    auto st = ThriftRpcHelper::rpc<FrontendServiceClient>(
354
0
            master_addr.hostname, master_addr.port,
355
0
            [&](FrontendServiceConnection& client) { client->getTabletReplicaInfos(resp, req); });
356
0
    if (!st) {
357
0
        return ResultError(Status::InternalError(
358
0
                "failed to get tablet replica infos, rpc error={}, tablet_id={}", st.to_string(),
359
0
                tablet_id));
360
0
    }
361
362
0
    auto it = resp.tablet_replica_infos.find(tablet_id);
363
0
    if (it == resp.tablet_replica_infos.end()) {
364
0
        return ResultError(Status::InternalError("replicas not found, tablet_id={}", tablet_id));
365
0
    }
366
0
    auto replicas = it->second;
367
0
    auto local_host = BackendOptions::get_localhost();
368
0
    bool include_local_host = false;
369
0
    DBUG_EXECUTE_IF("get_peer_replicas_address.enable_local_host", { include_local_host = true; });
370
0
    auto ret_view =
371
0
            replicas | std::views::filter([&local_host, include_local_host](const auto& replica) {
372
0
                return local_host.find(replica.host) == std::string::npos || include_local_host;
373
0
            }) |
374
0
            std::views::transform([](auto& replica) {
375
0
                return std::make_pair(std::move(replica.host), replica.brpc_port);
376
0
            });
377
0
    return std::vector(ret_view.begin(), ret_view.end());
378
0
}
379
380
Result<CaptureRowsetResult> BaseTablet::_remote_capture_rowsets(
381
0
        const Version& version_range) const {
382
0
    auto start_tm_us = MonotonicMicros();
383
0
    Defer defer {
384
0
            [&]() { g_remote_fetch_tablet_rowsets_latency << MonotonicMicros() - start_tm_us; }};
385
0
#ifndef BE_TEST
386
0
    auto maybe_be_addresses = get_peer_replicas_addresses(tablet_id());
387
#else
388
    Result<std::vector<std::pair<std::string, int32_t>>> maybe_be_addresses;
389
    TEST_SYNC_POINT_CALLBACK("get_peer_replicas_addresses", &maybe_be_addresses);
390
#endif
391
0
    DBUG_EXECUTE_IF("Tablet::_remote_get_rowsets_meta.inject_replica_address_fail",
392
0
                    { maybe_be_addresses = ResultError(Status::InternalError("inject failure")); });
393
0
    if (!maybe_be_addresses) {
394
0
        return ResultError(std::move(maybe_be_addresses.error()));
395
0
    }
396
0
    auto be_addresses = std::move(maybe_be_addresses.value());
397
0
    if (be_addresses.empty()) {
398
0
        LOG(WARNING) << "no peers replica for tablet=" << tablet_id();
399
0
        return ResultError(Status::InternalError("no replicas for tablet={}", tablet_id()));
400
0
    }
401
402
0
    auto cntl = std::make_shared<GetRowsetsCntl>();
403
0
    cntl->tablet_id = tablet_id();
404
0
    cntl->req_addrs = std::move(be_addresses);
405
0
    cntl->version_range = version_range;
406
0
    bool need_delete_bitmap = need_read_delete_bitmap();
407
0
    CaptureRowsetResult result;
408
0
    if (need_delete_bitmap) {
409
0
        result.delete_bitmap =
410
0
                std::make_unique<DeleteBitmap>(_tablet_meta->delete_bitmap().snapshot());
411
0
        DeleteBitmapPB delete_bitmap_keys;
412
0
        auto keyset = result.delete_bitmap->delete_bitmap |
413
0
                      std::views::transform([](const auto& kv) { return kv.first; });
414
0
        for (const auto& key : keyset) {
415
0
            const auto& [rs_id, seg_id, version] = key;
416
0
            delete_bitmap_keys.mutable_rowset_ids()->Add(rs_id.to_string());
417
0
            delete_bitmap_keys.mutable_segment_ids()->Add(seg_id);
418
0
            delete_bitmap_keys.mutable_versions()->Add(version);
419
0
        }
420
0
        cntl->delete_bitmap_keys = std::move(delete_bitmap_keys);
421
0
    }
422
423
0
    RETURN_IF_ERROR_RESULT(cntl->start_req_bg());
424
0
    auto maybe_meta = cntl->wait_for_ret();
425
0
    if (!maybe_meta) {
426
0
        auto err = Status::InternalError(
427
0
                "tried to get rowsets from peer replicas and failed, "
428
0
                "reason={}",
429
0
                maybe_meta.error());
430
0
        return ResultError(std::move(err));
431
0
    }
432
433
0
    auto& remote_meta = maybe_meta.value();
434
0
    const auto& rs_metas = remote_meta.rowsets;
435
0
    for (const auto& rs_meta : rs_metas) {
436
0
        RowsetSharedPtr rs;
437
0
        auto st = RowsetFactory::create_rowset(_tablet_meta->tablet_schema(), {}, rs_meta, &rs);
438
0
        if (!st) {
439
0
            return ResultError(std::move(st));
440
0
        }
441
0
        result.rowsets.push_back(std::move(rs));
442
0
    }
443
0
    if (need_delete_bitmap) {
444
        DCHECK_NE(result.delete_bitmap, nullptr);
445
0
        result.delete_bitmap->merge(*remote_meta.delete_bitmap);
446
0
    }
447
0
    return result;
448
0
}
449
450
} // namespace doris