/
rabbit_variable_queue.erl
2758 lines (2508 loc) · 122 KB
/
rabbit_variable_queue.erl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
%% The contents of this file are subject to the Mozilla Public License
%% Version 1.1 (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.mozilla.org/MPL/
%%
%% Software distributed under the License is distributed on an "AS IS"
%% basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See
%% the License for the specific language governing rights and
%% limitations under the License.
%%
%% The Original Code is RabbitMQ.
%%
%% The Initial Developer of the Original Code is GoPivotal, Inc.
%% Copyright (c) 2007-2017 Pivotal Software, Inc. All rights reserved.
%%
-module(rabbit_variable_queue).
-export([init/3, terminate/2, delete_and_terminate/2, delete_crashed/1,
purge/1, purge_acks/1,
publish/6, publish_delivered/5,
batch_publish/4, batch_publish_delivered/4,
discard/4, drain_confirmed/1,
dropwhile/2, fetchwhile/4, fetch/2, drop/2, ack/2, requeue/2,
ackfold/4, fold/3, len/1, is_empty/1, depth/1,
set_ram_duration_target/2, ram_duration/1, needs_timeout/1, timeout/1,
handle_pre_hibernate/1, resume/1, msg_rates/1,
info/2, invoke/3, is_duplicate/2, set_queue_mode/2,
zip_msgs_and_acks/4, multiple_routing_keys/0]).
-export([start/1, stop/0]).
%% exported for testing only
-export([start_msg_store/2, stop_msg_store/0, init/6]).
%%----------------------------------------------------------------------------
%% Messages, and their position in the queue, can be in memory or on
%% disk, or both. Persistent messages will have both message and
%% position pushed to disk as soon as they arrive; transient messages
%% can be written to disk (and thus both types can be evicted from
%% memory) under memory pressure. The question of whether a message is
%% in RAM and whether it is persistent are orthogonal.
%%
%% Messages are persisted using the queue index and the message
%% store. Normally the queue index holds the position of the message
%% *within this queue* along with a couple of small bits of metadata,
%% while the message store holds the message itself (including headers
%% and other properties).
%%
%% However, as an optimisation, small messages can be embedded
%% directly in the queue index and bypass the message store
%% altogether.
%%
%% Definitions:
%%
%% alpha: this is a message where both the message itself, and its
%% position within the queue are held in RAM
%%
%% beta: this is a message where the message itself is only held on
%% disk (if persisted to the message store) but its position
%% within the queue is held in RAM.
%%
%% gamma: this is a message where the message itself is only held on
%% disk, but its position is both in RAM and on disk.
%%
%% delta: this is a collection of messages, represented by a single
%% term, where the messages and their position are only held on
%% disk.
%%
%% Note that for persistent messages, the message and its position
%% within the queue are always held on disk, *in addition* to being in
%% one of the above classifications.
%%
%% Also note that within this code, the term gamma seldom
%% appears. It's frequently the case that gammas are defined by betas
%% who have had their queue position recorded on disk.
%%
%% In general, messages move q1 -> q2 -> delta -> q3 -> q4, though
%% many of these steps are frequently skipped. q1 and q4 only hold
%% alphas, q2 and q3 hold both betas and gammas. When a message
%% arrives, its classification is determined. It is then added to the
%% rightmost appropriate queue.
%%
%% If a new message is determined to be a beta or gamma, q1 is
%% empty. If a new message is determined to be a delta, q1 and q2 are
%% empty (and actually q4 too).
%%
%% When removing messages from a queue, if q4 is empty then q3 is read
%% directly. If q3 becomes empty then the next segment's worth of
%% messages from delta are read into q3, reducing the size of
%% delta. If the queue is non empty, either q4 or q3 contain
%% entries. It is never permitted for delta to hold all the messages
%% in the queue.
%%
%% The duration indicated to us by the memory_monitor is used to
%% calculate, given our current ingress and egress rates, how many
%% messages we should hold in RAM (i.e. as alphas). We track the
%% ingress and egress rates for both messages and pending acks and
%% rates for both are considered when calculating the number of
%% messages to hold in RAM. When we need to push alphas to betas or
%% betas to gammas, we favour writing out messages that are further
%% from the head of the queue. This minimises writes to disk, as the
%% messages closer to the tail of the queue stay in the queue for
%% longer, thus do not need to be replaced as quickly by sending other
%% messages to disk.
%%
%% Whilst messages are pushed to disk and forgotten from RAM as soon
%% as requested by a new setting of the queue RAM duration, the
%% inverse is not true: we only load messages back into RAM as
%% demanded as the queue is read from. Thus only publishes to the
%% queue will take up available spare capacity.
%%
%% When we report our duration to the memory monitor, we calculate
%% average ingress and egress rates over the last two samples, and
%% then calculate our duration based on the sum of the ingress and
%% egress rates. More than two samples could be used, but it's a
%% balance between responding quickly enough to changes in
%% producers/consumers versus ignoring temporary blips. The problem
%% with temporary blips is that with just a few queues, they can have
%% substantial impact on the calculation of the average duration and
%% hence cause unnecessary I/O. Another alternative is to increase the
%% amqqueue_process:RAM_DURATION_UPDATE_PERIOD to beyond 5
%% seconds. However, that then runs the risk of being too slow to
%% inform the memory monitor of changes. Thus a 5 second interval,
%% plus a rolling average over the last two samples seems to work
%% well in practice.
%%
%% The sum of the ingress and egress rates is used because the egress
%% rate alone is not sufficient. Adding in the ingress rate means that
%% queues which are being flooded by messages are given more memory,
%% resulting in them being able to process the messages faster (by
%% doing less I/O, or at least deferring it) and thus helping keep
%% their mailboxes empty and thus the queue as a whole is more
%% responsive. If such a queue also has fast but previously idle
%% consumers, the consumer can then start to be driven as fast as it
%% can go, whereas if only egress rate was being used, the incoming
%% messages may have to be written to disk and then read back in,
%% resulting in the hard disk being a bottleneck in driving the
%% consumers. Generally, we want to give Rabbit every chance of
%% getting rid of messages as fast as possible and remaining
%% responsive, and using only the egress rate impacts that goal.
%%
%% Once the queue has more alphas than the target_ram_count, the
%% surplus must be converted to betas, if not gammas, if not rolled
%% into delta. The conditions under which these transitions occur
%% reflect the conflicting goals of minimising RAM cost per msg, and
%% minimising CPU cost per msg. Once the msg has become a beta, its
%% payload is no longer in RAM, thus a read from the msg_store must
%% occur before the msg can be delivered, but the RAM cost of a beta
%% is the same as a gamma, so converting a beta to gamma will not free
%% up any further RAM. To reduce the RAM cost further, the gamma must
%% be rolled into delta. Whilst recovering a beta or a gamma to an
%% alpha requires only one disk read (from the msg_store), recovering
%% a msg from within delta will require two reads (queue_index and
%% then msg_store). But delta has a near-0 per-msg RAM cost. So the
%% conflict is between using delta more, which will free up more
%% memory, but require additional CPU and disk ops, versus using delta
%% less and gammas and betas more, which will cost more memory, but
%% require fewer disk ops and less CPU overhead.
%%
%% In the case of a persistent msg published to a durable queue, the
%% msg is immediately written to the msg_store and queue_index. If
%% then additionally converted from an alpha, it'll immediately go to
%% a gamma (as it's already in queue_index), and cannot exist as a
%% beta. Thus a durable queue with a mixture of persistent and
%% transient msgs in it which has more messages than permitted by the
%% target_ram_count may contain an interspersed mixture of betas and
%% gammas in q2 and q3.
%%
%% There is then a ratio that controls how many betas and gammas there
%% can be. This is based on the target_ram_count and thus expresses
%% the fact that as the number of permitted alphas in the queue falls,
%% so should the number of betas and gammas fall (i.e. delta
%% grows). If q2 and q3 contain more than the permitted number of
%% betas and gammas, then the surplus are forcibly converted to gammas
%% (as necessary) and then rolled into delta. The ratio is that
%% delta/(betas+gammas+delta) equals
%% (betas+gammas+delta)/(target_ram_count+betas+gammas+delta). I.e. as
%% the target_ram_count shrinks to 0, so must betas and gammas.
%%
%% The conversion of betas to deltas is done if there are at least
%% ?IO_BATCH_SIZE betas in q2 & q3. This value should not be too small,
%% otherwise the frequent operations on the queues of q2 and q3 will not be
%% effectively amortised (switching the direction of queue access defeats
%% amortisation). Note that there is a natural upper bound due to credit_flow
%% limits on the alpha to beta conversion.
%%
%% The conversion from alphas to betas is chunked due to the
%% credit_flow limits of the msg_store. This further smooths the
%% effects of changes to the target_ram_count and ensures the queue
%% remains responsive even when there is a large amount of IO work to
%% do. The 'resume' callback is utilised to ensure that conversions
%% are done as promptly as possible whilst ensuring the queue remains
%% responsive.
%%
%% In the queue we keep track of both messages that are pending
%% delivery and messages that are pending acks. In the event of a
%% queue purge, we only need to load qi segments if the queue has
%% elements in deltas (i.e. it came under significant memory
%% pressure). In the event of a queue deletion, in addition to the
%% preceding, by keeping track of pending acks in RAM, we do not need
%% to search through qi segments looking for messages that are yet to
%% be acknowledged.
%%
%% Pending acks are recorded in memory by storing the message itself.
%% If the message has been sent to disk, we do not store the message
%% content. During memory reduction, pending acks containing message
%% content have that content removed and the corresponding messages
%% are pushed out to disk.
%%
%% Messages from pending acks are returned to q4, q3 and delta during
%% requeue, based on the limits of seq_id contained in each. Requeued
%% messages retain their original seq_id, maintaining order
%% when requeued.
%%
%% The order in which alphas are pushed to betas and pending acks
%% are pushed to disk is determined dynamically. We always prefer to
%% push messages for the source (alphas or acks) that is growing the
%% fastest (with growth measured as avg. ingress - avg. egress).
%%
%% Notes on Clean Shutdown
%% (This documents behaviour in variable_queue, queue_index and
%% msg_store.)
%%
%% In order to try to achieve as fast a start-up as possible, if a
%% clean shutdown occurs, we try to save out state to disk to reduce
%% work on startup. In the msg_store this takes the form of the
%% index_module's state, plus the file_summary ets table, and client
%% refs. In the VQ, this takes the form of the count of persistent
%% messages in the queue and references into the msg_stores. The
%% queue_index adds to these terms the details of its segments and
%% stores the terms in the queue directory.
%%
%% Two message stores are used. One is created for persistent messages
%% to durable queues that must survive restarts, and the other is used
%% for all other messages that just happen to need to be written to
%% disk. On start up we can therefore nuke the transient message
%% store, and be sure that the messages in the persistent store are
%% all that we need.
%%
%% The references to the msg_stores are there so that the msg_store
%% knows to only trust its saved state if all of the queues it was
%% previously talking to come up cleanly. Likewise, the queues
%% themselves (esp queue_index) skips work in init if all the queues
%% and msg_store were shutdown cleanly. This gives both good speed
%% improvements and also robustness so that if anything possibly went
%% wrong in shutdown (or there was subsequent manual tampering), all
%% messages and queues that can be recovered are recovered, safely.
%%
%% To delete transient messages lazily, the variable_queue, on
%% startup, stores the next_seq_id reported by the queue_index as the
%% transient_threshold. From that point on, whenever it's reading a
%% message off disk via the queue_index, if the seq_id is below this
%% threshold and the message is transient then it drops the message
%% (the message itself won't exist on disk because it would have been
%% stored in the transient msg_store which would have had its saved
%% state nuked on startup). This avoids the expensive operation of
%% scanning the entire queue on startup in order to delete transient
%% messages that were only pushed to disk to save memory.
%%
%%----------------------------------------------------------------------------
-behaviour(rabbit_backing_queue).
-record(vqstate,
{ q1,
q2,
delta,
q3,
q4,
next_seq_id,
ram_pending_ack, %% msgs using store, still in RAM
disk_pending_ack, %% msgs in store, paged out
qi_pending_ack, %% msgs using qi, *can't* be paged out
index_state,
msg_store_clients,
durable,
transient_threshold,
qi_embed_msgs_below,
len, %% w/o unacked
bytes, %% w/o unacked
unacked_bytes,
persistent_count, %% w unacked
persistent_bytes, %% w unacked
delta_transient_bytes, %%
target_ram_count,
ram_msg_count, %% w/o unacked
ram_msg_count_prev,
ram_ack_count_prev,
ram_bytes, %% w unacked
out_counter,
in_counter,
rates,
msgs_on_disk,
msg_indices_on_disk,
unconfirmed,
confirmed,
ack_out_counter,
ack_in_counter,
%% Unlike the other counters these two do not feed into
%% #rates{} and get reset
disk_read_count,
disk_write_count,
io_batch_size,
%% default queue or lazy queue
mode,
%% number of reduce_memory_usage executions, once it
%% reaches a threshold the queue will manually trigger a runtime GC
%% see: maybe_execute_gc/1
memory_reduction_run_count
}).
-record(rates, { in, out, ack_in, ack_out, timestamp }).
-record(msg_status,
{ seq_id,
msg_id,
msg,
is_persistent,
is_delivered,
msg_in_store,
index_on_disk,
persist_to,
msg_props
}).
-record(delta,
{ start_seq_id, %% start_seq_id is inclusive
count,
transient,
end_seq_id %% end_seq_id is exclusive
}).
-define(HEADER_GUESS_SIZE, 100). %% see determine_persist_to/2
-define(PERSISTENT_MSG_STORE, msg_store_persistent).
-define(TRANSIENT_MSG_STORE, msg_store_transient).
-define(QUEUE, lqueue).
-include("rabbit.hrl").
-include("rabbit_framing.hrl").
%%----------------------------------------------------------------------------
-rabbit_upgrade({multiple_routing_keys, local, []}).
-type seq_id() :: non_neg_integer().
-type rates() :: #rates { in :: float(),
out :: float(),
ack_in :: float(),
ack_out :: float(),
timestamp :: rabbit_types:timestamp()}.
-type delta() :: #delta { start_seq_id :: non_neg_integer(),
count :: non_neg_integer(),
end_seq_id :: non_neg_integer() }.
%% The compiler (rightfully) complains that ack() and state() are
%% unused. For this reason we duplicate a -spec from
%% rabbit_backing_queue with the only intent being to remove
%% warnings. The problem here is that we can't parameterise the BQ
%% behaviour by these two types as we would like to. We still leave
%% these here for documentation purposes.
-type ack() :: seq_id().
-type state() :: #vqstate {
q1 :: ?QUEUE:?QUEUE(),
q2 :: ?QUEUE:?QUEUE(),
delta :: delta(),
q3 :: ?QUEUE:?QUEUE(),
q4 :: ?QUEUE:?QUEUE(),
next_seq_id :: seq_id(),
ram_pending_ack :: gb_trees:tree(),
disk_pending_ack :: gb_trees:tree(),
qi_pending_ack :: gb_trees:tree(),
index_state :: any(),
msg_store_clients :: 'undefined' | {{any(), binary()},
{any(), binary()}},
durable :: boolean(),
transient_threshold :: non_neg_integer(),
qi_embed_msgs_below :: non_neg_integer(),
len :: non_neg_integer(),
bytes :: non_neg_integer(),
unacked_bytes :: non_neg_integer(),
persistent_count :: non_neg_integer(),
persistent_bytes :: non_neg_integer(),
target_ram_count :: non_neg_integer() | 'infinity',
ram_msg_count :: non_neg_integer(),
ram_msg_count_prev :: non_neg_integer(),
ram_ack_count_prev :: non_neg_integer(),
ram_bytes :: non_neg_integer(),
out_counter :: non_neg_integer(),
in_counter :: non_neg_integer(),
rates :: rates(),
msgs_on_disk :: ?GB_SET_TYPE(),
msg_indices_on_disk :: ?GB_SET_TYPE(),
unconfirmed :: ?GB_SET_TYPE(),
confirmed :: ?GB_SET_TYPE(),
ack_out_counter :: non_neg_integer(),
ack_in_counter :: non_neg_integer(),
disk_read_count :: non_neg_integer(),
disk_write_count :: non_neg_integer(),
io_batch_size :: pos_integer(),
mode :: 'default' | 'lazy',
memory_reduction_run_count :: non_neg_integer()}.
%% Duplicated from rabbit_backing_queue
-spec ack([ack()], state()) -> {[rabbit_guid:guid()], state()}.
-spec multiple_routing_keys() -> 'ok'.
-define(BLANK_DELTA, #delta { start_seq_id = undefined,
count = 0,
transient = 0,
end_seq_id = undefined }).
-define(BLANK_DELTA_PATTERN(Z), #delta { start_seq_id = Z,
count = 0,
transient = 0,
end_seq_id = Z }).
-define(MICROS_PER_SECOND, 1000000.0).
%% We're sampling every 5s for RAM duration; a half life that is of
%% the same order of magnitude is probably about right.
-define(RATE_AVG_HALF_LIFE, 5.0).
%% We will recalculate the #rates{} every time we get asked for our
%% RAM duration, or every N messages published, whichever is
%% sooner. We do this since the priority calculations in
%% rabbit_amqqueue_process need fairly fresh rates.
-define(MSGS_PER_RATE_CALC, 100).
%% we define the garbage collector threshold
%% it needs to tune the `reduce_memory_use` calls. Thus, the garbage collection.
%% see: rabbitmq-server-973 and rabbitmq-server-964
-define(DEFAULT_EXPLICIT_GC_RUN_OP_THRESHOLD, 1000).
-define(EXPLICIT_GC_RUN_OP_THRESHOLD(Mode),
case get(explicit_gc_run_operation_threshold) of
undefined ->
Val = explicit_gc_run_operation_threshold_for_mode(Mode),
put(explicit_gc_run_operation_threshold, Val),
Val;
Val -> Val
end).
explicit_gc_run_operation_threshold_for_mode(Mode) ->
{Key, Fallback} = case Mode of
lazy -> {lazy_queue_explicit_gc_run_operation_threshold,
?DEFAULT_EXPLICIT_GC_RUN_OP_THRESHOLD};
_ -> {queue_explicit_gc_run_operation_threshold,
?DEFAULT_EXPLICIT_GC_RUN_OP_THRESHOLD}
end,
rabbit_misc:get_env(rabbit, Key, Fallback).
%%----------------------------------------------------------------------------
%% Public API
%%----------------------------------------------------------------------------
start(DurableQueues) ->
{AllTerms, StartFunState} = rabbit_queue_index:start(DurableQueues),
start_msg_store(
[Ref || Terms <- AllTerms,
Terms /= non_clean_shutdown,
begin
Ref = proplists:get_value(persistent_ref, Terms),
Ref =/= undefined
end],
StartFunState),
{ok, AllTerms}.
stop() ->
ok = stop_msg_store(),
ok = rabbit_queue_index:stop().
start_msg_store(Refs, StartFunState) ->
ok = rabbit_sup:start_child(?TRANSIENT_MSG_STORE, rabbit_msg_store,
[?TRANSIENT_MSG_STORE, rabbit_mnesia:dir(),
undefined, {fun (ok) -> finished end, ok}]),
ok = rabbit_sup:start_child(?PERSISTENT_MSG_STORE, rabbit_msg_store,
[?PERSISTENT_MSG_STORE, rabbit_mnesia:dir(),
Refs, StartFunState]).
stop_msg_store() ->
ok = rabbit_sup:stop_child(?PERSISTENT_MSG_STORE),
ok = rabbit_sup:stop_child(?TRANSIENT_MSG_STORE).
init(Queue, Recover, Callback) ->
init(
Queue, Recover, Callback,
fun (MsgIds, ActionTaken) ->
msgs_written_to_disk(Callback, MsgIds, ActionTaken)
end,
fun (MsgIds) -> msg_indices_written_to_disk(Callback, MsgIds) end,
fun (MsgIds) -> msgs_and_indices_written_to_disk(Callback, MsgIds) end).
init(#amqqueue { name = QueueName, durable = IsDurable }, new,
AsyncCallback, MsgOnDiskFun, MsgIdxOnDiskFun, MsgAndIdxOnDiskFun) ->
IndexState = rabbit_queue_index:init(QueueName,
MsgIdxOnDiskFun, MsgAndIdxOnDiskFun),
init(IsDurable, IndexState, 0, 0, [],
case IsDurable of
true -> msg_store_client_init(?PERSISTENT_MSG_STORE,
MsgOnDiskFun, AsyncCallback);
false -> undefined
end,
msg_store_client_init(?TRANSIENT_MSG_STORE, undefined, AsyncCallback));
%% We can be recovering a transient queue if it crashed
init(#amqqueue { name = QueueName, durable = IsDurable }, Terms,
AsyncCallback, MsgOnDiskFun, MsgIdxOnDiskFun, MsgAndIdxOnDiskFun) ->
{PRef, RecoveryTerms} = process_recovery_terms(Terms),
{PersistentClient, ContainsCheckFun} =
case IsDurable of
true -> C = msg_store_client_init(?PERSISTENT_MSG_STORE, PRef,
MsgOnDiskFun, AsyncCallback),
{C, fun (MsgId) when is_binary(MsgId) ->
rabbit_msg_store:contains(MsgId, C);
(#basic_message{is_persistent = Persistent}) ->
Persistent
end};
false -> {undefined, fun(_MsgId) -> false end}
end,
TransientClient = msg_store_client_init(?TRANSIENT_MSG_STORE,
undefined, AsyncCallback),
{DeltaCount, DeltaBytes, IndexState} =
rabbit_queue_index:recover(
QueueName, RecoveryTerms,
rabbit_msg_store:successfully_recovered_state(?PERSISTENT_MSG_STORE),
ContainsCheckFun, MsgIdxOnDiskFun, MsgAndIdxOnDiskFun),
init(IsDurable, IndexState, DeltaCount, DeltaBytes, RecoveryTerms,
PersistentClient, TransientClient).
process_recovery_terms(Terms=non_clean_shutdown) ->
{rabbit_guid:gen(), Terms};
process_recovery_terms(Terms) ->
case proplists:get_value(persistent_ref, Terms) of
undefined -> {rabbit_guid:gen(), []};
PRef -> {PRef, Terms}
end.
terminate(_Reason, State) ->
State1 = #vqstate { persistent_count = PCount,
persistent_bytes = PBytes,
index_state = IndexState,
msg_store_clients = {MSCStateP, MSCStateT} } =
purge_pending_ack(true, State),
PRef = case MSCStateP of
undefined -> undefined;
_ -> ok = rabbit_msg_store:client_terminate(MSCStateP),
rabbit_msg_store:client_ref(MSCStateP)
end,
ok = rabbit_msg_store:client_delete_and_terminate(MSCStateT),
Terms = [{persistent_ref, PRef},
{persistent_count, PCount},
{persistent_bytes, PBytes}],
a(State1 #vqstate { index_state = rabbit_queue_index:terminate(
Terms, IndexState),
msg_store_clients = undefined }).
%% the only difference between purge and delete is that delete also
%% needs to delete everything that's been delivered and not ack'd.
delete_and_terminate(_Reason, State) ->
%% Normally when we purge messages we interact with the qi by
%% issues delivers and acks for every purged message. In this case
%% we don't need to do that, so we just delete the qi.
State1 = purge_and_index_reset(State),
State2 = #vqstate { msg_store_clients = {MSCStateP, MSCStateT} } =
purge_pending_ack_delete_and_terminate(State1),
case MSCStateP of
undefined -> ok;
_ -> rabbit_msg_store:client_delete_and_terminate(MSCStateP)
end,
rabbit_msg_store:client_delete_and_terminate(MSCStateT),
a(State2 #vqstate { msg_store_clients = undefined }).
delete_crashed(#amqqueue{name = QName}) ->
ok = rabbit_queue_index:erase(QName).
purge(State = #vqstate { len = Len }) ->
case is_pending_ack_empty(State) and is_unconfirmed_empty(State) of
true ->
{Len, purge_and_index_reset(State)};
false ->
{Len, purge_when_pending_acks(State)}
end.
purge_acks(State) -> a(purge_pending_ack(false, State)).
publish(Msg, MsgProps, IsDelivered, ChPid, Flow, State) ->
State1 =
publish1(Msg, MsgProps, IsDelivered, ChPid, Flow,
fun maybe_write_to_disk/4,
State),
a(maybe_reduce_memory_use(maybe_update_rates(State1))).
batch_publish(Publishes, ChPid, Flow, State) ->
{ChPid, Flow, State1} =
lists:foldl(fun batch_publish1/2, {ChPid, Flow, State}, Publishes),
State2 = ui(State1),
a(maybe_reduce_memory_use(maybe_update_rates(State2))).
publish_delivered(Msg, MsgProps, ChPid, Flow, State) ->
{SeqId, State1} =
publish_delivered1(Msg, MsgProps, ChPid, Flow,
fun maybe_write_to_disk/4,
State),
{SeqId, a(maybe_reduce_memory_use(maybe_update_rates(State1)))}.
batch_publish_delivered(Publishes, ChPid, Flow, State) ->
{ChPid, Flow, SeqIds, State1} =
lists:foldl(fun batch_publish_delivered1/2,
{ChPid, Flow, [], State}, Publishes),
State2 = ui(State1),
{lists:reverse(SeqIds), a(maybe_reduce_memory_use(maybe_update_rates(State2)))}.
discard(_MsgId, _ChPid, _Flow, State) -> State.
drain_confirmed(State = #vqstate { confirmed = C }) ->
case gb_sets:is_empty(C) of
true -> {[], State}; %% common case
false -> {gb_sets:to_list(C), State #vqstate {
confirmed = gb_sets:new() }}
end.
dropwhile(Pred, State) ->
{MsgProps, State1} =
remove_by_predicate(Pred, State),
{MsgProps, a(State1)}.
fetchwhile(Pred, Fun, Acc, State) ->
{MsgProps, Acc1, State1} =
fetch_by_predicate(Pred, Fun, Acc, State),
{MsgProps, Acc1, a(State1)}.
fetch(AckRequired, State) ->
case queue_out(State) of
{empty, State1} ->
{empty, a(State1)};
{{value, MsgStatus}, State1} ->
%% it is possible that the message wasn't read from disk
%% at this point, so read it in.
{Msg, State2} = read_msg(MsgStatus, State1),
{AckTag, State3} = remove(AckRequired, MsgStatus, State2),
{{Msg, MsgStatus#msg_status.is_delivered, AckTag}, a(State3)}
end.
drop(AckRequired, State) ->
case queue_out(State) of
{empty, State1} ->
{empty, a(State1)};
{{value, MsgStatus}, State1} ->
{AckTag, State2} = remove(AckRequired, MsgStatus, State1),
{{MsgStatus#msg_status.msg_id, AckTag}, a(State2)}
end.
ack([], State) ->
{[], State};
%% optimisation: this head is essentially a partial evaluation of the
%% general case below, for the single-ack case.
ack([SeqId], State) ->
case remove_pending_ack(true, SeqId, State) of
{none, _} ->
State;
{#msg_status { msg_id = MsgId,
is_persistent = IsPersistent,
msg_in_store = MsgInStore,
index_on_disk = IndexOnDisk },
State1 = #vqstate { index_state = IndexState,
msg_store_clients = MSCState,
ack_out_counter = AckOutCount }} ->
IndexState1 = case IndexOnDisk of
true -> rabbit_queue_index:ack([SeqId], IndexState);
false -> IndexState
end,
case MsgInStore of
true -> ok = msg_store_remove(MSCState, IsPersistent, [MsgId]);
false -> ok
end,
{[MsgId],
a(State1 #vqstate { index_state = IndexState1,
ack_out_counter = AckOutCount + 1 })}
end;
ack(AckTags, State) ->
{{IndexOnDiskSeqIds, MsgIdsByStore, AllMsgIds},
State1 = #vqstate { index_state = IndexState,
msg_store_clients = MSCState,
ack_out_counter = AckOutCount }} =
lists:foldl(
fun (SeqId, {Acc, State2}) ->
case remove_pending_ack(true, SeqId, State2) of
{none, _} ->
{Acc, State2};
{MsgStatus, State3} ->
{accumulate_ack(MsgStatus, Acc), State3}
end
end, {accumulate_ack_init(), State}, AckTags),
IndexState1 = rabbit_queue_index:ack(IndexOnDiskSeqIds, IndexState),
remove_msgs_by_id(MsgIdsByStore, MSCState),
{lists:reverse(AllMsgIds),
a(State1 #vqstate { index_state = IndexState1,
ack_out_counter = AckOutCount + length(AckTags) })}.
requeue(AckTags, #vqstate { mode = default,
delta = Delta,
q3 = Q3,
q4 = Q4,
in_counter = InCounter,
len = Len } = State) ->
{SeqIds, Q4a, MsgIds, State1} = queue_merge(lists:sort(AckTags), Q4, [],
beta_limit(Q3),
fun publish_alpha/2, State),
{SeqIds1, Q3a, MsgIds1, State2} = queue_merge(SeqIds, Q3, MsgIds,
delta_limit(Delta),
fun publish_beta/2, State1),
{Delta1, MsgIds2, State3} = delta_merge(SeqIds1, Delta, MsgIds1,
State2),
MsgCount = length(MsgIds2),
{MsgIds2, a(maybe_reduce_memory_use(
maybe_update_rates(ui(
State3 #vqstate { delta = Delta1,
q3 = Q3a,
q4 = Q4a,
in_counter = InCounter + MsgCount,
len = Len + MsgCount }))))};
requeue(AckTags, #vqstate { mode = lazy,
delta = Delta,
q3 = Q3,
in_counter = InCounter,
len = Len } = State) ->
{SeqIds, Q3a, MsgIds, State1} = queue_merge(lists:sort(AckTags), Q3, [],
delta_limit(Delta),
fun publish_beta/2, State),
{Delta1, MsgIds1, State2} = delta_merge(SeqIds, Delta, MsgIds,
State1),
MsgCount = length(MsgIds1),
{MsgIds1, a(maybe_reduce_memory_use(
maybe_update_rates(ui(
State2 #vqstate { delta = Delta1,
q3 = Q3a,
in_counter = InCounter + MsgCount,
len = Len + MsgCount }))))}.
ackfold(MsgFun, Acc, State, AckTags) ->
{AccN, StateN} =
lists:foldl(fun(SeqId, {Acc0, State0}) ->
MsgStatus = lookup_pending_ack(SeqId, State0),
{Msg, State1} = read_msg(MsgStatus, State0),
{MsgFun(Msg, SeqId, Acc0), State1}
end, {Acc, State}, AckTags),
{AccN, a(StateN)}.
fold(Fun, Acc, State = #vqstate{index_state = IndexState}) ->
{Its, IndexState1} = lists:foldl(fun inext/2, {[], IndexState},
[msg_iterator(State),
disk_ack_iterator(State),
ram_ack_iterator(State),
qi_ack_iterator(State)]),
ifold(Fun, Acc, Its, State#vqstate{index_state = IndexState1}).
len(#vqstate { len = Len }) -> Len.
is_empty(State) -> 0 == len(State).
depth(State) ->
len(State) + count_pending_acks(State).
set_ram_duration_target(
DurationTarget, State = #vqstate {
rates = #rates { in = AvgIngressRate,
out = AvgEgressRate,
ack_in = AvgAckIngressRate,
ack_out = AvgAckEgressRate },
target_ram_count = TargetRamCount }) ->
Rate =
AvgEgressRate + AvgIngressRate + AvgAckEgressRate + AvgAckIngressRate,
TargetRamCount1 =
case DurationTarget of
infinity -> infinity;
_ -> trunc(DurationTarget * Rate) %% msgs = sec * msgs/sec
end,
State1 = State #vqstate { target_ram_count = TargetRamCount1 },
a(case TargetRamCount1 == infinity orelse
(TargetRamCount =/= infinity andalso
TargetRamCount1 >= TargetRamCount) of
true -> State1;
false -> reduce_memory_use(State1)
end).
maybe_update_rates(State = #vqstate{ in_counter = InCount,
out_counter = OutCount })
when InCount + OutCount > ?MSGS_PER_RATE_CALC ->
update_rates(State);
maybe_update_rates(State) ->
State.
update_rates(State = #vqstate{ in_counter = InCount,
out_counter = OutCount,
ack_in_counter = AckInCount,
ack_out_counter = AckOutCount,
rates = #rates{ in = InRate,
out = OutRate,
ack_in = AckInRate,
ack_out = AckOutRate,
timestamp = TS }}) ->
Now = time_compat:monotonic_time(),
Rates = #rates { in = update_rate(Now, TS, InCount, InRate),
out = update_rate(Now, TS, OutCount, OutRate),
ack_in = update_rate(Now, TS, AckInCount, AckInRate),
ack_out = update_rate(Now, TS, AckOutCount, AckOutRate),
timestamp = Now },
State#vqstate{ in_counter = 0,
out_counter = 0,
ack_in_counter = 0,
ack_out_counter = 0,
rates = Rates }.
update_rate(Now, TS, Count, Rate) ->
Time = time_compat:convert_time_unit(Now - TS, native, micro_seconds) /
?MICROS_PER_SECOND,
if
Time == 0 -> Rate;
true -> rabbit_misc:moving_average(Time, ?RATE_AVG_HALF_LIFE,
Count / Time, Rate)
end.
ram_duration(State) ->
State1 = #vqstate { rates = #rates { in = AvgIngressRate,
out = AvgEgressRate,
ack_in = AvgAckIngressRate,
ack_out = AvgAckEgressRate },
ram_msg_count = RamMsgCount,
ram_msg_count_prev = RamMsgCountPrev,
ram_pending_ack = RPA,
qi_pending_ack = QPA,
ram_ack_count_prev = RamAckCountPrev } =
update_rates(State),
RamAckCount = gb_trees:size(RPA) + gb_trees:size(QPA),
Duration = %% msgs+acks / (msgs+acks/sec) == sec
case lists:all(fun (X) -> X < 0.01 end,
[AvgEgressRate, AvgIngressRate,
AvgAckEgressRate, AvgAckIngressRate]) of
true -> infinity;
false -> (RamMsgCountPrev + RamMsgCount +
RamAckCount + RamAckCountPrev) /
(4 * (AvgEgressRate + AvgIngressRate +
AvgAckEgressRate + AvgAckIngressRate))
end,
{Duration, State1}.
needs_timeout(#vqstate { index_state = IndexState }) ->
case rabbit_queue_index:needs_sync(IndexState) of
confirms -> timed;
other -> idle;
false -> false
end.
timeout(State = #vqstate { index_state = IndexState }) ->
State #vqstate { index_state = rabbit_queue_index:sync(IndexState) }.
handle_pre_hibernate(State = #vqstate { index_state = IndexState }) ->
State #vqstate { index_state = rabbit_queue_index:flush(IndexState) }.
resume(State) -> a(reduce_memory_use(State)).
msg_rates(#vqstate { rates = #rates { in = AvgIngressRate,
out = AvgEgressRate } }) ->
{AvgIngressRate, AvgEgressRate}.
info(messages_ready_ram, #vqstate{ram_msg_count = RamMsgCount}) ->
RamMsgCount;
info(messages_unacknowledged_ram, #vqstate{ram_pending_ack = RPA,
qi_pending_ack = QPA}) ->
gb_trees:size(RPA) + gb_trees:size(QPA);
info(messages_ram, State) ->
info(messages_ready_ram, State) + info(messages_unacknowledged_ram, State);
info(messages_persistent, #vqstate{persistent_count = PersistentCount}) ->
PersistentCount;
info(messages_paged_out, #vqstate{delta = #delta{transient = Count}}) ->
Count;
info(message_bytes, #vqstate{bytes = Bytes,
unacked_bytes = UBytes}) ->
Bytes + UBytes;
info(message_bytes_ready, #vqstate{bytes = Bytes}) ->
Bytes;
info(message_bytes_unacknowledged, #vqstate{unacked_bytes = UBytes}) ->
UBytes;
info(message_bytes_ram, #vqstate{ram_bytes = RamBytes}) ->
RamBytes;
info(message_bytes_persistent, #vqstate{persistent_bytes = PersistentBytes}) ->
PersistentBytes;
info(message_bytes_paged_out, #vqstate{delta_transient_bytes = PagedOutBytes}) ->
PagedOutBytes;
info(head_message_timestamp, #vqstate{
q3 = Q3,
q4 = Q4,
ram_pending_ack = RPA,
qi_pending_ack = QPA}) ->
head_message_timestamp(Q3, Q4, RPA, QPA);
info(disk_reads, #vqstate{disk_read_count = Count}) ->
Count;
info(disk_writes, #vqstate{disk_write_count = Count}) ->
Count;
info(backing_queue_status, #vqstate {
q1 = Q1, q2 = Q2, delta = Delta, q3 = Q3, q4 = Q4,
mode = Mode,
len = Len,
target_ram_count = TargetRamCount,
next_seq_id = NextSeqId,
rates = #rates { in = AvgIngressRate,
out = AvgEgressRate,
ack_in = AvgAckIngressRate,
ack_out = AvgAckEgressRate }}) ->
[ {mode , Mode},
{q1 , ?QUEUE:len(Q1)},
{q2 , ?QUEUE:len(Q2)},
{delta , Delta},
{q3 , ?QUEUE:len(Q3)},
{q4 , ?QUEUE:len(Q4)},
{len , Len},
{target_ram_count , TargetRamCount},
{next_seq_id , NextSeqId},
{avg_ingress_rate , AvgIngressRate},
{avg_egress_rate , AvgEgressRate},
{avg_ack_ingress_rate, AvgAckIngressRate},
{avg_ack_egress_rate , AvgAckEgressRate} ];
info(Item, _) ->
throw({bad_argument, Item}).
invoke(?MODULE, Fun, State) -> Fun(?MODULE, State);
invoke( _, _, State) -> State.
is_duplicate(_Msg, State) -> {false, State}.
set_queue_mode(Mode, State = #vqstate { mode = Mode }) ->
State;
set_queue_mode(lazy, State = #vqstate {
target_ram_count = TargetRamCount }) ->
%% To become a lazy queue we need to page everything to disk first.
State1 = convert_to_lazy(State),
%% restore the original target_ram_count
a(State1 #vqstate { mode = lazy, target_ram_count = TargetRamCount });
set_queue_mode(default, State) ->
%% becoming a default queue means loading messages from disk like
%% when a queue is recovered.
a(maybe_deltas_to_betas(State #vqstate { mode = default }));
set_queue_mode(_, State) ->
State.
zip_msgs_and_acks(Msgs, AckTags, Accumulator, _State) ->
lists:foldl(fun ({{#basic_message{ id = Id }, _Props}, AckTag}, Acc) ->
[{Id, AckTag} | Acc]
end, Accumulator, lists:zip(Msgs, AckTags)).
convert_to_lazy(State) ->
State1 = #vqstate { delta = Delta, q3 = Q3, len = Len } =
set_ram_duration_target(0, State),
case Delta#delta.count + ?QUEUE:len(Q3) == Len of
true ->
State1;
false ->
%% When pushing messages to disk, we might have been
%% blocked by the msg_store, so we need to see if we have
%% to wait for more credit, and then keep paging messages.
%%
%% The amqqueue_process could have taken care of this, but
%% between the time it receives the bump_credit msg and
%% calls BQ:resume to keep paging messages to disk, some
%% other request may arrive to the BQ which at this moment
%% is not in a proper state for a lazy BQ (unless all
%% messages have been paged to disk already).
wait_for_msg_store_credit(),
convert_to_lazy(resume(State1))
end.
wait_for_msg_store_credit() ->
case credit_flow:blocked() of
true -> receive
{bump_credit, Msg} ->
credit_flow:handle_bump_msg(Msg)
end;
false -> ok
end.
%% Get the Timestamp property of the first msg, if present. This is
%% the one with the oldest timestamp among the heads of the pending
%% acks and unread queues. We can't check disk_pending_acks as these
%% are paged out - we assume some will soon be paged in rather than
%% forcing it to happen. Pending ack msgs are included as they are