nsqd: set memoryMsgChan as nil when --mem-queue-size=0#1159
Conversation
mreiferson
changed the title
|
Thanks for your warmly, patient answer, but in my opinion it should give the user an option to have a chance to save all the messages on the disk-queue. Do you think so? @mreiferson |
|
This isn't really enough to "guarantee" that, though. The disk segments are not necessarily synced after every message, so the message might still be buffered in the page cache and not written to the physical disk, though you could set But more importantly, messages are taken out of the disk queue while they are in flight, and exist in the in-memory in-flight pqueue. If the server crashes hard at this point, they will probably be lost. They may still exist in some disk segment file, but the metadata will say that they are in an already-consumed part of that disk segment file, and they won't be re-played. (And they might have been in a segment file that was finished and cleaned-up). To get what you're going for, you need something like #625 |
|
Without considering the network factors,It's very usefull when i need to keep message order. |
|
OK ... I'm convinced this change is a good idea. There are still serious limitations to the "disk-queue-only" mode which make me skeptical of its value, but anyone who configured |
do not use unbuffered chan for in-memory queue when size=0 because user intended all messages to go through backend queue (disk)
|
I took the liberty of changing the code and comment style. Local testing with |
In b3b29b7 / nsqio#1159 unbuffered memoryMsgChan were replaced with nil chan for both Topic and Channel (this only happens if mem-queue-size=0). This inadvertently made deferred publish never work (with mem-queue-size=0), because some metadata is lost when messages go through the "backend" disk queue, but previously messages could go immediately through the topic's unbuffered chan with their metadata intact, to then be added to each channel's deferred pqueue. This partial revert brings this part back. Granted, mem-queue-size=0 never worked very well, in many respects. For this functionality, at low message publish rates, it seems there was rarely a problem. But at high message publish rates it was always likely that some messages would hit the topic disk queue and lose the deferred time metadata. This potential problem remains. In fact, the motivation for fully disabling the memoryMsgQueue for mem-queue-size=0 was to avoid excessively shuffling message order, which would happen if some messages jump instantly through the memory-queue while others take the longer way through the disk-queue. That will be likely again. But the users who noticed this probably did not use publish-deferred functionality! Additional fixes would be appropriate, but this is a quick-fix to restore previous useful behavior for mem-queue-size=0.
In b3b29b7 / nsqio#1159 unbuffered memoryMsgChan were replaced with nil chan for both Topic and Channel if mem-queue-size=0. This inadvertently made deferred publish never work (with mem-queue-size=0), because some metadata is lost when messages go through the "backend" disk queue, instead of immediately going through the topic memory chan to the channels' deferred pqueues. The motivation for fully disabling the memoryMsgQueue for mem-queue-size=0 was to avoid excessively shuffling message order, which would happen if some messages jump instantly through the memory-queue while others take the longer way through the disk-queue. So, only allow using the memory queue in this case if really needed, for deferred messages, or for ephemeral topic or channel which just lose all messages if they all go through the no-op backend queue. Granted, mem-queue-size=0 never worked very well, in many respects.
In b3b29b7 / nsqio#1159 unbuffered memoryMsgChan were replaced with nil chan for both Topic and Channel if mem-queue-size=0. This inadvertently made deferred publish never work (with mem-queue-size=0), because some metadata is lost when messages go through the "backend" disk queue, instead of immediately going through the topic memory chan to the channels' deferred pqueues. The motivation for fully disabling the memoryMsgQueue for mem-queue-size=0 was to avoid excessively shuffling message order, which would happen if some messages jump instantly through the memory-queue while others take the longer way through the disk-queue. So, only allow using the memory queue in this case if really needed, for deferred messages, or for ephemeral topic or channel which just lose all messages if they all go through the no-op backend queue. Granted, mem-queue-size=0 never worked very well, in many respects.
4 participants
Note: generally, we intend to save all the message in the disk-queue when set
-mem-queue-size=0option, butmake(chan *Message, 0)just create a unbuffered chan which can also send/receive message as well.the Go channel spec dictates that nil channel operations (read or write) in a select are skipped, we set memoryMsgChan as not nil noly when mem-queue-size > 0 as shown in diskqueue.go