diff options
Diffstat (limited to 'src/rabbit_variable_queue.erl')
| -rw-r--r-- | src/rabbit_variable_queue.erl | 604 |
1 files changed, 413 insertions, 191 deletions
diff --git a/src/rabbit_variable_queue.erl b/src/rabbit_variable_queue.erl index 69d62fde..0db51165 100644 --- a/src/rabbit_variable_queue.erl +++ b/src/rabbit_variable_queue.erl @@ -42,7 +42,7 @@ -export([start/1, stop/0]). %% exported for testing only --export([start_msg_store/2, stop_msg_store/0]). +-export([start_msg_store/2, stop_msg_store/0, init/5]). %%---------------------------------------------------------------------------- %% Definitions: @@ -89,12 +89,14 @@ %% %% 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. 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. +%% messages we should hold in RAM. 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 @@ -156,7 +158,7 @@ %% The conversion from alphas to betas is also chunked, but only to %% ensure no more than ?IO_BATCH_SIZE alphas are converted to betas at %% any one time. This further smooths the effects of changes to the -%% target_ram_msg_count and ensures the queue remains responsive +%% target_ram_count and ensures the queue remains responsive %% even when there is a large amount of IO work to do. The %% idle_timeout callback is utilised to ensure that conversions are %% done as promptly as possible whilst ensuring the queue remains @@ -168,6 +170,29 @@ %% the latter) are both cheap and do require any scanning through qi %% segments. %% +%% Pending acks are recorded in memory either as the tuple {SeqId, +%% Guid, MsgProps} (tuple-form) or as the message itself (message- +%% form). Acks for persistent messages are always stored in the tuple- +%% form. Acks for transient messages are also stored in tuple-form if +%% the message has been sent to disk as part of the memory reduction +%% process. For transient messages that haven't already been written +%% to disk, acks are stored in message-form. +%% +%% During memory reduction, acks stored in message-form are converted +%% to tuple-form, and the corresponding messages are pushed out to +%% disk. +%% +%% The order in which alphas are pushed to betas and message-form 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). In +%% each round of memory reduction a chunk of messages at most +%% ?IO_BATCH_SIZE in size is allocated to be pushed to disk. The +%% fastest growing source will be reduced by as much of this chunk as +%% possible. If there is any remaining allocation in the chunk after +%% the first source has been reduced to zero, the second source will +%% be reduced by as much of the remaining chunk as possible. +%% %% Notes on Clean Shutdown %% (This documents behaviour in variable_queue, queue_index and %% msg_store.) @@ -220,6 +245,8 @@ q4, next_seq_id, pending_ack, + pending_ack_index, + ram_ack_index, index_state, msg_store_clients, on_sync, @@ -229,14 +256,21 @@ len, persistent_count, - target_ram_msg_count, + target_ram_count, ram_msg_count, ram_msg_count_prev, + ram_ack_count_prev, ram_index_count, out_counter, in_counter, - rates - }). + rates, + msgs_on_disk, + msg_indices_on_disk, + unconfirmed, + ack_out_counter, + ack_in_counter, + ack_rates + }). -record(rates, { egress, ingress, avg_egress, avg_ingress, timestamp }). @@ -299,30 +333,37 @@ funs :: [fun (() -> any())] }). -type(state() :: #vqstate { - q1 :: queue(), - q2 :: bpqueue:bpqueue(), - delta :: delta(), - q3 :: bpqueue:bpqueue(), - q4 :: queue(), - next_seq_id :: seq_id(), - pending_ack :: dict(), - index_state :: any(), - msg_store_clients :: 'undefined' | {{any(), binary()}, + q1 :: queue(), + q2 :: bpqueue:bpqueue(), + delta :: delta(), + q3 :: bpqueue:bpqueue(), + q4 :: queue(), + next_seq_id :: seq_id(), + pending_ack :: dict(), + ram_ack_index :: gb_tree(), + index_state :: any(), + msg_store_clients :: 'undefined' | {{any(), binary()}, {any(), binary()}}, - on_sync :: sync(), - durable :: boolean(), - - len :: non_neg_integer(), - persistent_count :: non_neg_integer(), - - transient_threshold :: non_neg_integer(), - target_ram_msg_count :: non_neg_integer() | 'infinity', - ram_msg_count :: non_neg_integer(), - ram_msg_count_prev :: non_neg_integer(), - ram_index_count :: non_neg_integer(), - out_counter :: non_neg_integer(), - in_counter :: non_neg_integer(), - rates :: rates() }). + on_sync :: sync(), + durable :: boolean(), + + len :: non_neg_integer(), + persistent_count :: non_neg_integer(), + + transient_threshold :: non_neg_integer(), + target_ram_count :: non_neg_integer() | 'infinity', + ram_msg_count :: non_neg_integer(), + ram_msg_count_prev :: non_neg_integer(), + ram_index_count :: non_neg_integer(), + out_counter :: non_neg_integer(), + in_counter :: non_neg_integer(), + rates :: rates(), + msgs_on_disk :: gb_set(), + msg_indices_on_disk :: gb_set(), + unconfirmed :: gb_set(), + ack_out_counter :: non_neg_integer(), + ack_in_counter :: non_neg_integer(), + ack_rates :: rates() }). -include("rabbit_backing_queue_spec.hrl"). @@ -368,16 +409,23 @@ stop_msg_store() -> ok = rabbit_sup:stop_child(?PERSISTENT_MSG_STORE), ok = rabbit_sup:stop_child(?TRANSIENT_MSG_STORE). -init(QueueName, IsDurable, false) -> - IndexState = rabbit_queue_index:init(QueueName), +init(QueueName, IsDurable, Recover) -> + Self = self(), + init(QueueName, IsDurable, Recover, + fun (Guids) -> msgs_written_to_disk(Self, Guids) end, + fun (Guids) -> msg_indices_written_to_disk(Self, Guids) end). + +init(QueueName, IsDurable, false, MsgOnDiskFun, MsgIdxOnDiskFun) -> + IndexState = rabbit_queue_index:init(QueueName, MsgIdxOnDiskFun), init(IsDurable, IndexState, 0, [], case IsDurable of - true -> msg_store_client_init(?PERSISTENT_MSG_STORE); + true -> msg_store_client_init(?PERSISTENT_MSG_STORE, + MsgOnDiskFun); false -> undefined end, - msg_store_client_init(?TRANSIENT_MSG_STORE)); + msg_store_client_init(?TRANSIENT_MSG_STORE, undefined)); -init(QueueName, true, true) -> +init(QueueName, true, true, MsgOnDiskFun, MsgIdxOnDiskFun) -> Terms = rabbit_queue_index:shutdown_terms(QueueName), {PRef, TRef, Terms1} = case [persistent_ref, transient_ref] -- proplists:get_keys(Terms) of @@ -387,16 +435,17 @@ init(QueueName, true, true) -> _ -> {rabbit_guid:guid(), rabbit_guid:guid(), []} end, PersistentClient = rabbit_msg_store:client_init(?PERSISTENT_MSG_STORE, - PRef), + PRef, MsgOnDiskFun), TransientClient = rabbit_msg_store:client_init(?TRANSIENT_MSG_STORE, - TRef), + TRef, undefined), {DeltaCount, IndexState} = rabbit_queue_index:recover( QueueName, Terms1, rabbit_msg_store:successfully_recovered_state(?PERSISTENT_MSG_STORE), fun (Guid) -> rabbit_msg_store:contains(Guid, PersistentClient) - end), + end, + MsgIdxOnDiskFun), init(true, IndexState, DeltaCount, Terms1, PersistentClient, TransientClient). @@ -472,26 +521,30 @@ publish(Msg, MsgProps, State) -> publish_delivered(false, _Msg, _MsgProps, State = #vqstate { len = 0 }) -> {blank_ack, a(State)}; -publish_delivered(true, Msg = #basic_message { is_persistent = IsPersistent }, - MsgProps, +publish_delivered(true, Msg = #basic_message { is_persistent = IsPersistent, + guid = Guid }, + MsgProps = #message_properties { + needs_confirming = NeedsConfirming }, State = #vqstate { len = 0, next_seq_id = SeqId, out_counter = OutCount, in_counter = InCount, persistent_count = PCount, - pending_ack = PA, - durable = IsDurable }) -> + durable = IsDurable, + unconfirmed = Unconfirmed }) -> IsPersistent1 = IsDurable andalso IsPersistent, MsgStatus = (msg_status(IsPersistent1, SeqId, Msg, MsgProps)) #msg_status { is_delivered = true }, {MsgStatus1, State1} = maybe_write_to_disk(false, false, MsgStatus, State), - PA1 = record_pending_ack(m(MsgStatus1), PA), + State2 = record_pending_ack(m(MsgStatus1), State1), PCount1 = PCount + one_if(IsPersistent1), - {SeqId, a(State1 #vqstate { next_seq_id = SeqId + 1, - out_counter = OutCount + 1, - in_counter = InCount + 1, - persistent_count = PCount1, - pending_ack = PA1 })}. + Unconfirmed1 = gb_sets_maybe_insert(NeedsConfirming, Guid, Unconfirmed), + {SeqId, a(reduce_memory_use( + State2 #vqstate { next_seq_id = SeqId + 1, + out_counter = OutCount + 1, + in_counter = InCount + 1, + persistent_count = PCount1, + unconfirmed = Unconfirmed1 }))}. dropwhile(Pred, State) -> {_OkOrEmpty, State1} = dropwhile1(Pred, State), @@ -561,8 +614,7 @@ internal_fetch(AckRequired, MsgStatus = #msg_status { index_state = IndexState, msg_store_clients = MSCState, len = Len, - persistent_count = PCount, - pending_ack = PA }) -> + persistent_count = PCount }) -> %% 1. Mark it delivered if necessary IndexState1 = maybe_write_delivered( IndexOnDisk andalso not IsDelivered, @@ -582,12 +634,12 @@ internal_fetch(AckRequired, MsgStatus = #msg_status { end, %% 3. If an ack is required, add something sensible to PA - {AckTag, PA1} = case AckRequired of - true -> PA2 = record_pending_ack( - MsgStatus #msg_status { - is_delivered = true }, PA), - {SeqId, PA2}; - false -> {blank_ack, PA} + {AckTag, State1} = case AckRequired of + true -> StateN = record_pending_ack( + MsgStatus #msg_status { + is_delivered = true }, State), + {SeqId, StateN}; + false -> {blank_ack, State} end, PCount1 = PCount - one_if(IsPersistent andalso not AckRequired), @@ -595,17 +647,22 @@ internal_fetch(AckRequired, MsgStatus = #msg_status { RamMsgCount1 = RamMsgCount - one_if(Msg =/= undefined), {{Msg, IsDelivered, AckTag, Len1}, - a(State #vqstate { ram_msg_count = RamMsgCount1, - out_counter = OutCount + 1, - index_state = IndexState2, - len = Len1, - persistent_count = PCount1, - pending_ack = PA1 })}. + a(State1 #vqstate { ram_msg_count = RamMsgCount1, + out_counter = OutCount + 1, + index_state = IndexState2, + len = Len1, + persistent_count = PCount1 })}. ack(AckTags, State) -> - a(ack(fun msg_store_remove/3, - fun (_AckEntry, State1) -> State1 end, - AckTags, State)). + {Guids, State1} = + ack(fun msg_store_remove/3, + fun ({_IsPersistent, Guid, _MsgProps}, State1) -> + remove_confirms(gb_sets:singleton(Guid), State1); + (#msg_status{msg = #basic_message{guid = Guid}}, State1) -> + remove_confirms(gb_sets:singleton(Guid), State1) + end, + AckTags, State), + {Guids, a(State1)}. tx_publish(Txn, Msg = #basic_message { is_persistent = IsPersistent }, MsgProps, State = #vqstate { durable = IsDurable, @@ -655,7 +712,7 @@ tx_commit(Txn, Fun, MsgPropsFun, end)}. requeue(AckTags, MsgPropsFun, State) -> - a(reduce_memory_use( + {_Guids, State1} = ack(fun msg_store_release/3, fun (#msg_status { msg = Msg, msg_props = MsgProps }, State1) -> {_SeqId, State2} = publish(Msg, MsgPropsFun(MsgProps), @@ -670,48 +727,69 @@ requeue(AckTags, MsgPropsFun, State) -> true, true, State2), State3 end, - AckTags, State))). + AckTags, State), + a(reduce_memory_use(State1)). len(#vqstate { len = Len }) -> Len. is_empty(State) -> 0 == len(State). -set_ram_duration_target(DurationTarget, - State = #vqstate { - rates = #rates { avg_egress = AvgEgressRate, - avg_ingress = AvgIngressRate }, - target_ram_msg_count = TargetRamMsgCount }) -> - Rate = AvgEgressRate + AvgIngressRate, - TargetRamMsgCount1 = +set_ram_duration_target( + DurationTarget, State = #vqstate { + rates = #rates { avg_egress = AvgEgressRate, + avg_ingress = AvgIngressRate }, + ack_rates = #rates { avg_egress = AvgAckEgressRate, + avg_ingress = AvgAckIngressRate }, + 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_msg_count = TargetRamMsgCount1 }, - a(case TargetRamMsgCount1 == infinity orelse - (TargetRamMsgCount =/= infinity andalso - TargetRamMsgCount1 >= TargetRamMsgCount) of + 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). ram_duration(State = #vqstate { - rates = #rates { egress = Egress, - ingress = Ingress, - timestamp = Timestamp } = Rates, + rates = #rates { timestamp = Timestamp, + egress = Egress, + ingress = Ingress } = Rates, + ack_rates = #rates { timestamp = AckTimestamp, + egress = AckEgress, + ingress = AckIngress } = ARates, in_counter = InCount, out_counter = OutCount, + ack_in_counter = AckInCount, + ack_out_counter = AckOutCount, ram_msg_count = RamMsgCount, - ram_msg_count_prev = RamMsgCountPrev }) -> + ram_msg_count_prev = RamMsgCountPrev, + ram_ack_index = RamAckIndex, + ram_ack_count_prev = RamAckCountPrev }) -> Now = now(), {AvgEgressRate, Egress1} = update_rate(Now, Timestamp, OutCount, Egress), {AvgIngressRate, Ingress1} = update_rate(Now, Timestamp, InCount, Ingress), - Duration = %% msgs / (msgs/sec) == sec - case AvgEgressRate == 0 andalso AvgIngressRate == 0 of + {AvgAckEgressRate, AckEgress1} = + update_rate(Now, AckTimestamp, AckOutCount, AckEgress), + {AvgAckIngressRate, AckIngress1} = + update_rate(Now, AckTimestamp, AckInCount, AckIngress), + + RamAckCount = gb_trees:size(RamAckIndex), + + Duration = %% msgs+acks / (msgs+acks/sec) == sec + case AvgEgressRate == 0 andalso AvgIngressRate == 0 andalso + AvgAckEgressRate == 0 andalso AvgAckIngressRate == 0 of true -> infinity; - false -> (RamMsgCountPrev + RamMsgCount) / - (2 * (AvgEgressRate + AvgIngressRate)) + false -> (RamMsgCountPrev + RamMsgCount + + RamAckCount + RamAckCountPrev) / + (4 * (AvgEgressRate + AvgIngressRate + + AvgAckEgressRate + AvgAckIngressRate)) end, {Duration, State #vqstate { @@ -721,14 +799,24 @@ ram_duration(State = #vqstate { avg_egress = AvgEgressRate, avg_ingress = AvgIngressRate, timestamp = Now }, + ack_rates = ARates #rates { + egress = AckEgress1, + ingress = AckIngress1, + avg_egress = AvgAckEgressRate, + avg_ingress = AvgAckIngressRate, + timestamp = Now }, in_counter = 0, out_counter = 0, - ram_msg_count_prev = RamMsgCount }}. + ack_in_counter = 0, + ack_out_counter = 0, + ram_msg_count_prev = RamMsgCount, + ram_ack_count_prev = RamAckCount }}. needs_idle_timeout(State = #vqstate { on_sync = ?BLANK_SYNC }) -> - {Res, _State} = reduce_memory_use(fun (_Quota, State1) -> State1 end, + {Res, _State} = reduce_memory_use(fun (_Quota, State1) -> {0, State1} end, fun (_Quota, State1) -> State1 end, fun (State1) -> State1 end, + fun (_Quota, State1) -> {0, State1} end, State), Res; needs_idle_timeout(_State) -> @@ -739,33 +827,39 @@ idle_timeout(State) -> a(reduce_memory_use(tx_commit_index(State))). handle_pre_hibernate(State = #vqstate { index_state = IndexState }) -> State #vqstate { index_state = rabbit_queue_index:flush(IndexState) }. -status(#vqstate { q1 = Q1, q2 = Q2, delta = Delta, q3 = Q3, q4 = Q4, - len = Len, - pending_ack = PA, - on_sync = #sync { funs = From }, - target_ram_msg_count = TargetRamMsgCount, - ram_msg_count = RamMsgCount, - ram_index_count = RamIndexCount, - next_seq_id = NextSeqId, - persistent_count = PersistentCount, - rates = #rates { - avg_egress = AvgEgressRate, - avg_ingress = AvgIngressRate } }) -> - [ {q1 , queue:len(Q1)}, - {q2 , bpqueue:len(Q2)}, - {delta , Delta}, - {q3 , bpqueue:len(Q3)}, - {q4 , queue:len(Q4)}, - {len , Len}, - {pending_acks , dict:size(PA)}, - {outstanding_txns , length(From)}, - {target_ram_msg_count , TargetRamMsgCount}, - {ram_msg_count , RamMsgCount}, - {ram_index_count , RamIndexCount}, - {next_seq_id , NextSeqId}, - {persistent_count , PersistentCount}, - {avg_egress_rate , AvgEgressRate}, - {avg_ingress_rate , AvgIngressRate} ]. +status(#vqstate { + q1 = Q1, q2 = Q2, delta = Delta, q3 = Q3, q4 = Q4, + len = Len, + pending_ack = PA, + ram_ack_index = RAI, + on_sync = #sync { funs = From }, + target_ram_count = TargetRamCount, + ram_msg_count = RamMsgCount, + ram_index_count = RamIndexCount, + next_seq_id = NextSeqId, + persistent_count = PersistentCount, + rates = #rates { avg_egress = AvgEgressRate, + avg_ingress = AvgIngressRate }, + ack_rates = #rates { avg_egress = AvgAckEgressRate, + avg_ingress = AvgAckIngressRate } }) -> + [ {q1 , queue:len(Q1)}, + {q2 , bpqueue:len(Q2)}, + {delta , Delta}, + {q3 , bpqueue:len(Q3)}, + {q4 , queue:len(Q4)}, + {len , Len}, + {pending_acks , dict:size(PA)}, + {outstanding_txns , length(From)}, + {target_ram_count , TargetRamCount}, + {ram_msg_count , RamMsgCount}, + {ram_ack_count , gb_trees:size(RAI)}, + {ram_index_count , RamIndexCount}, + {next_seq_id , NextSeqId}, + {persistent_count , PersistentCount}, + {avg_ingress_rate , AvgIngressRate}, + {avg_egress_rate , AvgEgressRate}, + {avg_ack_ingress_rate, AvgAckIngressRate}, + {avg_ack_egress_rate , AvgAckEgressRate} ]. %%---------------------------------------------------------------------------- %% Minor helpers @@ -811,6 +905,10 @@ one_if(false) -> 0. cons_if(true, E, L) -> [E | L]; cons_if(false, _E, L) -> L. +gb_sets_maybe_insert(false, _Val, Set) -> Set; +%% when requeueing, we re-add a guid to the unconfimred set +gb_sets_maybe_insert(true, Val, Set) -> gb_sets:add(Val, Set). + msg_status(IsPersistent, SeqId, Msg = #basic_message { guid = Guid }, MsgProps) -> #msg_status { seq_id = SeqId, guid = Guid, msg = Msg, @@ -832,8 +930,8 @@ with_immutable_msg_store_state(MSCState, IsPersistent, Fun) -> end), Res. -msg_store_client_init(MsgStore) -> - rabbit_msg_store:client_init(MsgStore, rabbit_guid:guid()). +msg_store_client_init(MsgStore, MsgOnDiskFun) -> + rabbit_msg_store:client_init(MsgStore, rabbit_guid:guid(), MsgOnDiskFun). msg_store_write(MSCState, IsPersistent, Guid, Msg) -> with_immutable_msg_store_state( @@ -955,35 +1053,46 @@ init(IsDurable, IndexState, DeltaCount, Terms, end, Now = now(), State = #vqstate { - q1 = queue:new(), - q2 = bpqueue:new(), - delta = Delta, - q3 = bpqueue:new(), - q4 = queue:new(), - next_seq_id = NextSeqId, - pending_ack = dict:new(), - index_state = IndexState1, - msg_store_clients = {PersistentClient, TransientClient}, - on_sync = ?BLANK_SYNC, - durable = IsDurable, - transient_threshold = NextSeqId, - - len = DeltaCount1, - persistent_count = DeltaCount1, - - target_ram_msg_count = infinity, - ram_msg_count = 0, - ram_msg_count_prev = 0, - ram_index_count = 0, - out_counter = 0, - in_counter = 0, - rates = #rates { egress = {Now, 0}, - ingress = {Now, DeltaCount1}, - avg_egress = 0.0, - avg_ingress = 0.0, - timestamp = Now } }, + q1 = queue:new(), + q2 = bpqueue:new(), + delta = Delta, + q3 = bpqueue:new(), + q4 = queue:new(), + next_seq_id = NextSeqId, + pending_ack = dict:new(), + ram_ack_index = gb_trees:empty(), + index_state = IndexState1, + msg_store_clients = {PersistentClient, TransientClient}, + on_sync = ?BLANK_SYNC, + durable = IsDurable, + transient_threshold = NextSeqId, + + len = DeltaCount1, + persistent_count = DeltaCount1, + + target_ram_count = infinity, + ram_msg_count = 0, + ram_msg_count_prev = 0, + ram_ack_count_prev = 0, + ram_index_count = 0, + out_counter = 0, + in_counter = 0, + rates = blank_rate(Now, DeltaCount1), + msgs_on_disk = gb_sets:new(), + msg_indices_on_disk = gb_sets:new(), + unconfirmed = gb_sets:new(), + ack_out_counter = 0, + ack_in_counter = 0, + ack_rates = blank_rate(Now, 0) }, a(maybe_deltas_to_betas(State)). +blank_rate(Timestamp, IngressLength) -> + #rates { egress = {Timestamp, 0}, + ingress = {Timestamp, IngressLength}, + avg_egress = 0.0, + avg_ingress = 0.0, + timestamp = Timestamp }. + msg_store_callback(PersistentGuids, Pubs, AckTags, Fun, MsgPropsFun) -> Self = self(), F = fun () -> rabbit_amqqueue:maybe_run_queue_via_backing_queue( @@ -1000,7 +1109,7 @@ msg_store_callback(PersistentGuids, Pubs, AckTags, Fun, MsgPropsFun) -> end. remove_persistent_messages(Guids) -> - PersistentClient = msg_store_client_init(?PERSISTENT_MSG_STORE), + PersistentClient = msg_store_client_init(?PERSISTENT_MSG_STORE, undefined), ok = rabbit_msg_store:remove(Guids, PersistentClient), rabbit_msg_store:client_delete_and_terminate(PersistentClient). @@ -1051,6 +1160,7 @@ tx_commit_index(State = #vqstate { on_sync = #sync { durable = IsDurable }) -> PAcks = lists:append(SPAcks), Acks = lists:append(SAcks), + {_Guids, NewState} = ack(Acks, State), Pubs = [{Msg, Fun(MsgProps)} || {Fun, PubsN} <- lists:reverse(SPubs), {Msg, MsgProps} <- lists:reverse(PubsN)], {SeqIds, State1 = #vqstate { index_state = IndexState }} = @@ -1062,7 +1172,7 @@ tx_commit_index(State = #vqstate { on_sync = #sync { {SeqId, State3} = publish(Msg, MsgProps, false, IsPersistent1, State2), {cons_if(IsPersistent1, SeqId, SeqIdsAcc), State3} - end, {PAcks, ack(Acks, State)}, Pubs), + end, {PAcks, NewState}, Pubs), IndexState1 = rabbit_queue_index:sync(SeqIds, IndexState), [ Fun() || Fun <- lists:reverse(SFuns) ], reduce_memory_use( @@ -1116,15 +1226,17 @@ sum_guids_by_store_to_len(LensByStore, GuidsByStore) -> %% Internal gubbins for publishing %%---------------------------------------------------------------------------- -publish(Msg = #basic_message { is_persistent = IsPersistent }, - MsgProps, IsDelivered, MsgOnDisk, +publish(Msg = #basic_message { is_persistent = IsPersistent, guid = Guid }, + MsgProps = #message_properties { needs_confirming = NeedsConfirming }, + IsDelivered, MsgOnDisk, State = #vqstate { q1 = Q1, q3 = Q3, q4 = Q4, next_seq_id = SeqId, len = Len, in_counter = InCount, persistent_count = PCount, durable = IsDurable, - ram_msg_count = RamMsgCount }) -> + ram_msg_count = RamMsgCount, + unconfirmed = Unconfirmed }) -> IsPersistent1 = IsDurable andalso IsPersistent, MsgStatus = (msg_status(IsPersistent1, SeqId, Msg, MsgProps)) #msg_status { is_delivered = IsDelivered, msg_on_disk = MsgOnDisk}, @@ -1134,11 +1246,13 @@ publish(Msg = #basic_message { is_persistent = IsPersistent }, true -> State1 #vqstate { q4 = queue:in(m(MsgStatus1), Q4) } end, PCount1 = PCount + one_if(IsPersistent1), + Unconfirmed1 = gb_sets_maybe_insert(NeedsConfirming, Guid, Unconfirmed), {SeqId, State2 #vqstate { next_seq_id = SeqId + 1, len = Len + 1, in_counter = InCount + 1, persistent_count = PCount1, - ram_msg_count = RamMsgCount + 1}}. + ram_msg_count = RamMsgCount + 1, + unconfirmed = Unconfirmed1 }}. maybe_write_msg_to_disk(_Force, MsgStatus = #msg_status { msg_on_disk = true }, _MSCState) -> @@ -1191,12 +1305,21 @@ record_pending_ack(#msg_status { seq_id = SeqId, guid = Guid, is_persistent = IsPersistent, msg_on_disk = MsgOnDisk, - msg_props = MsgProps } = MsgStatus, PA) -> - AckEntry = case MsgOnDisk of - true -> {IsPersistent, Guid, MsgProps}; - false -> MsgStatus - end, - dict:store(SeqId, AckEntry, PA). + msg_props = MsgProps } = MsgStatus, + State = #vqstate { pending_ack = PA, + ram_ack_index = RAI, + ack_in_counter = AckInCount}) -> + {AckEntry, RAI1} = + case MsgOnDisk of + true -> + {{IsPersistent, Guid, MsgProps}, RAI}; + false -> + {MsgStatus, gb_trees:insert(SeqId, Guid, RAI)} + end, + PA1 = dict:store(SeqId, AckEntry, PA), + State #vqstate { pending_ack = PA1, + ram_ack_index = RAI1, + ack_in_counter = AckInCount + 1}. remove_pending_ack(KeepPersistent, State = #vqstate { pending_ack = PA, @@ -1204,7 +1327,8 @@ remove_pending_ack(KeepPersistent, msg_store_clients = MSCState }) -> {SeqIds, GuidsByStore} = dict:fold(fun accumulate_ack/3, {[], orddict:new()}, PA), - State1 = State #vqstate { pending_ack = dict:new() }, + State1 = State #vqstate { pending_ack = dict:new(), + ram_ack_index = gb_trees:empty() }, case KeepPersistent of true -> case orddict:find(false, GuidsByStore) of error -> State1; @@ -1221,27 +1345,35 @@ remove_pending_ack(KeepPersistent, end. ack(_MsgStoreFun, _Fun, [], State) -> - State; + {[], State}; ack(MsgStoreFun, Fun, AckTags, State) -> {{SeqIds, GuidsByStore}, State1 = #vqstate { index_state = IndexState, msg_store_clients = MSCState, - persistent_count = PCount }} = + persistent_count = PCount, + ack_out_counter = AckOutCount }} = lists:foldl( - fun (SeqId, {Acc, State2 = #vqstate { pending_ack = PA }}) -> + fun (SeqId, {Acc, State2 = #vqstate { pending_ack = PA, + ram_ack_index = RAI }}) -> AckEntry = dict:fetch(SeqId, PA), {accumulate_ack(SeqId, AckEntry, Acc), Fun(AckEntry, State2 #vqstate { - pending_ack = dict:erase(SeqId, PA) })} + pending_ack = dict:erase(SeqId, PA), + ram_ack_index = + gb_trees:delete_any(SeqId, RAI)})} end, {{[], orddict:new()}, State}, AckTags), IndexState1 = rabbit_queue_index:ack(SeqIds, IndexState), - ok = orddict:fold(fun (IsPersistent, Guids, ok) -> - MsgStoreFun(MSCState, IsPersistent, Guids) - end, ok, GuidsByStore), + AckdGuids = lists:concat( + orddict:fold( + fun (IsPersistent, Guids, Gs) -> + MsgStoreFun(MSCState, IsPersistent, Guids), + [Guids | Gs] + end, [], GuidsByStore)), PCount1 = PCount - find_persistent_count(sum_guids_by_store_to_len( orddict:new(), GuidsByStore)), - State1 #vqstate { index_state = IndexState1, - persistent_count = PCount1 }. + {AckdGuids, State1 #vqstate { index_state = IndexState1, + persistent_count = PCount1, + ack_out_counter = AckOutCount + length(AckTags) }}. accumulate_ack(_SeqId, #msg_status { is_persistent = false, %% ASSERTIONS msg_on_disk = false, @@ -1258,6 +1390,44 @@ find_persistent_count(LensByStore) -> end. %%---------------------------------------------------------------------------- +%% Internal plumbing for confirms (aka publisher acks) +%%---------------------------------------------------------------------------- + +remove_confirms(GuidSet, State = #vqstate { msgs_on_disk = MOD, + msg_indices_on_disk = MIOD, + unconfirmed = UC }) -> + State #vqstate { msgs_on_disk = gb_sets:difference(MOD, GuidSet), + msg_indices_on_disk = gb_sets:difference(MIOD, GuidSet), + unconfirmed = gb_sets:difference(UC, GuidSet) }. + +msgs_confirmed(GuidSet, State) -> + {{confirm, gb_sets:to_list(GuidSet)}, remove_confirms(GuidSet, State)}. + +msgs_written_to_disk(QPid, GuidSet) -> + rabbit_amqqueue:maybe_run_queue_via_backing_queue_async( + QPid, fun(State = #vqstate { msgs_on_disk = MOD, + msg_indices_on_disk = MIOD, + unconfirmed = UC }) -> + msgs_confirmed(gb_sets:intersection(GuidSet, MIOD), + State #vqstate { + msgs_on_disk = + gb_sets:intersection( + gb_sets:union(MOD, GuidSet), UC) }) + end). + +msg_indices_written_to_disk(QPid, GuidSet) -> + rabbit_amqqueue:maybe_run_queue_via_backing_queue_async( + QPid, fun(State = #vqstate { msgs_on_disk = MOD, + msg_indices_on_disk = MIOD, + unconfirmed = UC }) -> + msgs_confirmed(gb_sets:intersection(GuidSet, MOD), + State #vqstate { + msg_indices_on_disk = + gb_sets:intersection( + gb_sets:union(MIOD, GuidSet), UC) }) + end). + +%%---------------------------------------------------------------------------- %% Phase changes %%---------------------------------------------------------------------------- @@ -1270,7 +1440,7 @@ find_persistent_count(LensByStore) -> %% though the conversion function for that is called as necessary. The %% reason is twofold. Firstly, this is safe because the conversion is %% only ever necessary just after a transition to a -%% target_ram_msg_count of zero or after an incremental alpha->beta +%% target_ram_count of zero or after an incremental alpha->beta %% conversion. In the former case the conversion is performed straight %% away (i.e. any betas present at the time are converted to deltas), %% and in the latter case the need for a conversion is flagged up @@ -1280,26 +1450,77 @@ find_persistent_count(LensByStore) -> %% one segment's worth of messages in q3 - and thus would risk %% perpetually reporting the need for a conversion when no such %% conversion is needed. That in turn could cause an infinite loop. -reduce_memory_use(AlphaBetaFun, BetaGammaFun, BetaDeltaFun, State) -> - {Reduce, State1} = case chunk_size(State #vqstate.ram_msg_count, - State #vqstate.target_ram_msg_count) of - 0 -> {false, State}; - S1 -> {true, AlphaBetaFun(S1, State)} - end, - case State1 #vqstate.target_ram_msg_count of - infinity -> {Reduce, State1}; - 0 -> {Reduce, BetaDeltaFun(State1)}; - _ -> case chunk_size(State1 #vqstate.ram_index_count, - permitted_ram_index_count(State1)) of - ?IO_BATCH_SIZE = S2 -> {true, BetaGammaFun(S2, State1)}; - _ -> {Reduce, State1} - end +reduce_memory_use(_AlphaBetaFun, _BetaGammaFun, _BetaDeltaFun, _AckFun, + State = #vqstate {target_ram_count = infinity}) -> + {false, State}; +reduce_memory_use(AlphaBetaFun, BetaGammaFun, BetaDeltaFun, AckFun, + State = #vqstate { + ram_ack_index = RamAckIndex, + ram_msg_count = RamMsgCount, + target_ram_count = TargetRamCount, + rates = #rates { avg_ingress = AvgIngress, + avg_egress = AvgEgress }, + ack_rates = #rates { avg_ingress = AvgAckIngress, + avg_egress = AvgAckEgress } + }) -> + + {Reduce, State1} = + case chunk_size(RamMsgCount + gb_trees:size(RamAckIndex), + TargetRamCount) of + 0 -> {false, State}; + %% Reduce memory of pending acks and alphas. The order is + %% determined based on which is growing faster. Whichever + %% comes second may very well get a quota of 0 if the + %% first manages to push out the max number of messages. + S1 -> {_, State2} = + lists:foldl(fun (ReduceFun, {QuotaN, StateN}) -> + ReduceFun(QuotaN, StateN) + end, + {S1, State}, + case (AvgAckIngress - AvgAckEgress) > + (AvgIngress - AvgEgress) of + true -> [AckFun, AlphaBetaFun]; + false -> [AlphaBetaFun, AckFun] + end), + {true, State2} + end, + + case State1 #vqstate.target_ram_count of + 0 -> {Reduce, BetaDeltaFun(State1)}; + _ -> case chunk_size(State1 #vqstate.ram_index_count, + permitted_ram_index_count(State1)) of + ?IO_BATCH_SIZE = S2 -> {true, BetaGammaFun(S2, State1)}; + _ -> {Reduce, State1} + end end. +limit_ram_acks(0, State) -> + {0, State}; +limit_ram_acks(Quota, State = #vqstate { pending_ack = PA, + ram_ack_index = RAI }) -> + case gb_trees:is_empty(RAI) of + true -> + {Quota, State}; + false -> + {SeqId, Guid, RAI1} = gb_trees:take_largest(RAI), + MsgStatus = #msg_status { + guid = Guid, %% ASSERTION + is_persistent = false, %% ASSERTION + msg_props = MsgProps } = dict:fetch(SeqId, PA), + {_, State1} = maybe_write_to_disk(true, false, MsgStatus, State), + limit_ram_acks(Quota - 1, + State1 #vqstate { + pending_ack = + dict:store(SeqId, {false, Guid, MsgProps}, PA), + ram_ack_index = RAI1 }) + end. + + reduce_memory_use(State) -> {_, State1} = reduce_memory_use(fun push_alphas_to_betas/2, fun limit_ram_index/2, fun push_betas_to_deltas/1, + fun limit_ram_acks/2, State), State1. @@ -1432,9 +1653,9 @@ maybe_deltas_to_betas(State = #vqstate { end. push_alphas_to_betas(Quota, State) -> - { Quota1, State1} = maybe_push_q1_to_betas(Quota, State), - {_Quota2, State2} = maybe_push_q4_to_betas(Quota1, State1), - State2. + {Quota1, State1} = maybe_push_q1_to_betas(Quota, State), + {Quota2, State2} = maybe_push_q4_to_betas(Quota1, State1), + {Quota2, State2}. maybe_push_q1_to_betas(Quota, State = #vqstate { q1 = Q1 }) -> maybe_push_alphas_to_betas( @@ -1460,10 +1681,11 @@ maybe_push_q4_to_betas(Quota, State = #vqstate { q4 = Q4 }) -> maybe_push_alphas_to_betas(_Generator, _Consumer, Quota, _Q, State = #vqstate { - ram_msg_count = RamMsgCount, - target_ram_msg_count = TargetRamMsgCount }) + ram_msg_count = RamMsgCount, + target_ram_count = TargetRamCount }) when Quota =:= 0 orelse - TargetRamMsgCount =:= infinity orelse TargetRamMsgCount >= RamMsgCount -> + TargetRamCount =:= infinity orelse + TargetRamCount >= RamMsgCount -> {Quota, State}; maybe_push_alphas_to_betas(Generator, Consumer, Quota, Q, State) -> case Generator(Q) of |