diff options
Diffstat (limited to 'doc/queues.html')
| -rw-r--r-- | doc/queues.html | 250 |
1 files changed, 243 insertions, 7 deletions
diff --git a/doc/queues.html b/doc/queues.html index 7f651a4c..3037ee42 100644 --- a/doc/queues.html +++ b/doc/queues.html @@ -22,7 +22,8 @@ forwarding to another host).</p> <h1>Queue Modes</h1> <p>Rsyslog supports different queue modes, some with submodes. Each of them has specific advantages and disadvantages. Selecting the right queue mode is quite -important when tuning rsyslogd.</p> +important when tuning rsyslogd. The queue mode (aka "type") is set via the "<i>$<object>QueueType</i>" +config directive.</p> <h2>Direct Queues</h2> <p>Direct queues are <b>non</b>-queueing queues. A queue in direct mode does neither queue nor bufer any of the queue elements but rather passes the element @@ -39,6 +40,8 @@ queueing happens.</p> the execution return code (sucess/failure) from the consumer to the producer. This, for example, is needed for the backup action logic. Consequently, backup actions require the to-be-checked action to use a "direct" mode queue.</p> +<p>To create a direct queue, use the "<i>$<object>QueueType Direct</i>" config +directive.</p> <h2>Disk Queues</h2> <p>Disk queues use disk drives for buffering. The important fact is that the always use the disk and do not buffer anything in memory. Thus, the queue is @@ -52,6 +55,12 @@ is mangled with the file structures). However, disk queues can be set to write bookkeeping information on checkpoints (every n records), so that this can be made ultra-reliable, too. If the checkpoint interval is set to one, no data can be lost, but the queue is exceptionally slow.</p> +<p>Each queue can be placed on a different disk for best performance and/or +isolation. This is currently selected by specifying different <i>$WorkDirectory</i> +config directives before the queue creation statement.</p> +<p>To create a disk queue, use the "<i>$<object>QueueType Disk</i>" config +directive. Checkpoint intervals can be specified via "<i>$<object>QueueCheckpointInterval</i>", +with 0 meaning no checkpoints. </p> <h2>In-Memory Queues</h2> <p>In-memory queue mode is what most people have on their mind when they think about computing queues. Here, the enqueued data elements are held in memory. @@ -59,20 +68,247 @@ Consequently, in-memory queues are very fast. But of course, they do not survive any program or operating system abort (what usually is tolerable and unlikely). Be sure to use an UPS if you use in-memory mode and your log data is important to you. Note that even in-memory queues may hold data for an infinite amount of -time if e.g. an output destination system is down and there is no reason to move +time when e.g. an output destination system is down and there is no reason to move the data out of memory (lying around in memory for an extended period of time is NOT a reason). Pure in-memory queues can't even store queue elements anywhere -else than in core memory.</p> -<p>If a disk queue name is defined for in-memory queues, they automatically +else than in core memory. </p> +<p>There exist two different in-memory queue modes: LinkedList and FixedArray. +Both are quite similar from the user's point of view, but utilize different +algorithms. </p> +<p>A FixedArray queue uses a fixed, pre-allocated array that holds pointers to +queue elements. The majority of space is taken up by the actual user data +elements, to which the pointers in the array point. The pointer array itself is +comparatively small. However, it has a certain memory footprint even if the +queue is empty. As there is no need to dynamically allocate any housekeeping +structures, FixedArray offers the best run time performance (uses the least CPU +cycle). FixedArray is best if there is a relatively low number of queue elements +expected and performance is desired. It is the default mode for the main message +queue (with a limit of 10,000 elements).</p> +<p>A LinkedList queue is quite the opposite. All housekeeping structures are +dynamically allocated (in a linked list, as its name implies). This requires +somewhat more runtime processing overhead, but ensures that memory is only +allocated in cases where it is needed. LinkedList queues are especially +well-suited for queues where only occasionally a than-high number of elements +need to be queued. A use case may be occasional message burst. Memory +permitting, it could be limited to e.g. 200,000 elements which would take up +only memory if in use. A FixedArray queue may have a too large static memory +footprint in such cases.</p> +<p><b>In general, it is advised to use LinkedList mode if in doubt</b>. The +processing overhead compared to FixedArray is low and may be outwaight by the +reduction in memory use. Paging in most-often-unused pointer array pages can be +much slower than dynamically allocating them.</p> +<p>To create an in-memory queue, use the "<i>$<object>QueueType LinkedList</i>" +or "<i>$<object>QueueType FixedArray</i>" config directive.</p> +<h3>Disk-Assisted Memory Queues</h3> +<p>If a disk queue name is defined for in-memory queues (via <i> +$<object>QueueFileName</i>), they automatically become "disk-assisted" (DA). In that mode, data is written to disk (and read -back) on a per-needed basis. Actually, the regular memory queue (called the +back) on an as-needed basis.</p> +<p>Actually, the regular memory queue (called the "primary queue") and a disk queue (called the "DA queue") work in tandem in this mode. Most importantly, the disk queue is activated if the primary queue is full or needs to be persisted on shutdown. Disk-assisted queues combine the -advantages of pure memory queues with those from pure disk queues. Under normal +advantages of pure memory queues with those of pure disk queues. Under normal operations, they are very fast and messages will never touch the disk. But if there is need to, an unlimited amount of messages can be buffered (actually -limited by free space) and data can be persisted between rsyslogd runs.</p> +limited by free disk space only) and data can be persisted between rsyslogd runs.</p> +<p>With a DA-queue, both disk-specific and in-memory specific configuration +parameters can be set. From the user's point of view, think of a DA queue like a +"super-queue" which does all within a single queue [from the code perspective, +there is some specific handling for this case, so it is actually much like a +single object].</p> +<p>DA queues are typically used to de-couple potentially long-running and +unreliable actions (to make them reliable). For example, it is recommened to use +a disk-assisted linked list in-memory queue in front of each database and "send +via tcp" action. Doing so makes these actions reliable and de-couples their +potential low execution speed from the rest of your rules (e.g. the local file +writes). There is a howto on <a href="rsyslog_high_database_rate.html">massive +database inserts</a> which nicely describes this use case. It may even be a good +read if you do not intend to use databases.</p> +<p>With DA queues, we do not simply write out everything to disk and then run as +a disk queue once the in-memory queue is full. A much smarter algorithm is used, +which involves a "high watermark" and a "low watermark". Both specify numbers of +queued items. If the queue size reaches high watermark elements, the queue +begins to write data elements to disk. It does so until it reaches the low water +mark elements. At this point, it stops writing until either high water mark is +reached again or the on-disk queue becomes empty, in which case the queue +reverts back to in-memory mode, only. While holding at the low watermark, new +elements are actually enqueued in memory. They are eventually written to disk, +but only if the high water mark is ever reached again. If it isn't, these items +never touch the disk. So even when a queue runs disk-assisted, there is +in-memory data present (this is a big difference to pure disk queues!).</p> +<p>This algorithm prevents unnecessary disk writes, but also leaves some +additional buffer space for message bursts. Remember that creating disk files +and writing to them is a lengthy operation. It is too lengthy to e.g. block +receiving UDP messages. Doing so would result in message loss. Thus, the queue +initiates DA mode, but still is able to receive messages and enqueue them - as +long as the maximum queue size is not reached. The number of elements between +the high water mark and the maximum queue size serves as this "emergency +buffer". Size it according to your needs, if traffic is very bursty you will +probably need a large buffer here. Keep in mind, though, that under normal +operations these queue elements will probably never be used. Setting the high +water mark too low will cause disk-assistance to be turned on more often than +actually needed.</p> +<p>The water marks can be set via the "<i>$<object>QueueHighWatermark</i>" and +"<i>$<object>QueueHighWatermark</i>" configuration file directives. Note that +these are actual numbers, not precentages. Be sure they make sense (also in +respect to "<i>$<object>QueueSize</i>"), as rsyslodg does currently not perform +any checks on the numbers provided. It is easy to screw up the system here (yes, +a feature enhancement request is filed ;)).</p> +<h1>Limiting the Queue Size</h1> +<p>All queues, including disk queues, have a limit of the number of elements +they can enqueue. This is set via the "<i>$<object>QueueSize</i>" config +parameter. Note that the size is specified in number of enqueued elements, not +their actual memory size. Memory size limits can not be set. A conservative +assumption is that a single syslog messages takes up 512 bytes on average +(in-memory, NOT on the wire, this *is* a difference).</p> +<p>Disk assisted queues are special in that they do <b>not</b> have any size +limit. The enqueue an unlimited amount of elements. To prevent running out of +space, disk and disk-assisted queues can be size-limited via the "<i>$<object>QueueMaxDiskSpace</i>" +configuration parameter. If it is not set, the limit is only available free +space (and reaching this limit is currently not very gracefully handled, so +avoid running into it!). If a limit is set, the queue can not grow larger than +it. Note, however, that the limit is approximate. The engine always writes +complete records. As such, it is possible that slightly more than the set limit +is used (usually less than 1k, given the average message size). Keeping strictly +on the limit would be a performance hurt, and thus the design decision was to +favour performance. If you don't like that policy, simply specify a slightly +lower limit (e.g. 999,999K instead of 1G).</p> +<p>In general, it is a good idea to limit the pysical disk space even if you +dedicate a whole disk to rsyslog. That way, you prevent it from running out of +space (future version will have an auto-size-limit logic, that then kicks in in +such situations).</p> +<h1>Worker Thread Pools</h1> +<p>Each queue (except in "direct" mode) has an associated pool of worker +threads. Worker threads carry out the action to be performed on the data +elements enqueued. As an actual sample, the main message queue's worker task is +to apply filter logic to each incoming message and enqueue them to the relevant +output queues (actions).</p> +<p>Worker threads are started and stopped on an as-needed bassis. On a system +without activity, there may be no worker at all running. One is automatically +started when a message comes in. Similarily, additional workers are started if +the queue grows above a specific size. The "<i>$<object>QueueWorkerThreadMinimumMessages</i>" +config parameter controls worker startup. If it is set to the minimum number of +elements that must be enqueued in order to justify a new worker startup. For +example, let's assume it is set to 100. As long as no more than 100 messages are +in the queue, a single worker will be used. When more than 100 messages arrive, +a new worker thread is automatically started. Similarily, a third worker will be +started when there are at least 300 messages, a forth when reaching 400 and so +on.</p> +<p>It, however, does not make sense to have too many worker threads running in +parall. Thus, the upper limit ca be set via "<i>$<object>QueueMaxWorkerThreads</i>". +If it, for example, is set to four, no more than four workers will ever be +started, no matter how many elements are enqueued. </p> +<p>Worker threads that have been started are kept running until an inactivity +timeout happens. The timeout can be set via "<i>$<object>QueueWorkerTimeoutShutdown</i>" +and is specified in milliseconds. If you do not like to keep the workers +running, simply set it to 0, which means immediate timeout and thus immediate +shutdown. But consider that creating threads involves some overhead, and this is +why we keep them running.</p> +<h2>Discarding Messages</h2> +<p>If the queue reaches the so called "discard watermark" (a number of queued +elements), less important messages can automatically be discarded. This is in an +effort to save queue space for more important messages, which you even less like +to loose. Please note that whenever there are more than "discard watermark" +messages, both newly incoming as well as already enqueued low-priority messages +are discarded. The algorithm discards messages newly coming in and those at the +front of the queue.</p> +<p>The discard watermark is a last resort setting. It should be set sufficiently +high, but low enough to allow for large message burst. Please note that it take +effect immediately and thus shows effect promptly - but that doesn't help if the +burst mainly consits of high-priority messages...</p> +<p>The discard watermark is set via the "<i>$<object>QueueDiscardMark</i>" +directive. The priority of messages to be discarded is set via "<i>$<object>QueueDiscardSeverity</i>". +Please note that as of now, this directive does not accept textual severity, so +a number according to the following table must be chosen (from RFC 3164):</p> +<pre> Numerical Severity + Code + + 0 Emergency: system is unusable + 1 Alert: action must be taken immediately + 2 Critical: critical conditions + 3 Error: error conditions + 4 Warning: warning conditions + 5 Notice: normal but significant condition + 6 Informational: informational messages + 7 Debug: debug-level messages</pre> +<p>Anything of the specified severity and (numerically) above it is discarded. +To turn message discarding off, simply specify the discard watermark to be +higher than the queue size.</p> +<p>An interesting application is with disk-assisted queues: if the discard +watermark is set lower than the high watermark, message discarding will start +before the queue becomes disk-assisted. This may be a good thing if you would +like to switch to disk-assisted mode only in cases where it is absolutelty +unavoidable and you prefer to discard less important messages first.</p> +<h1>Filled-Up Queues</h1> +<p>If the queue has either reached its configured maximum number of entries or +disk space, it is finally full. If so, rsyslogd throttles the data element +submitter. If that, for example, is a reliable input (TCP, local log socket), +that will slow down the message originator which is a good resulotion for this +scenario.</p> +<p>During throtteling, a disk-assisted queue continues to write to disk and +messages are also discarded based on severity as well as regular dequeuing and +processing continues. So chances are good the situation will be resolved by +simply throttling. Note, though, that throtteling is highly undesirable for +unreliable sources, like UDP message reception. So it is not a good thing to run +into throtteling mode at all.</p> +<p>We can not hold processing infenitely, not even when throtteling. For +example, throtteling the local log socket too long would cause the system at +whole come to a standstill. To prevent this, rsyslogd times out after a +configured period ("<i>$<object>QueueTimeoutEnqueue</i>", specified in +milliseconds) if no space becomes available. As a last resort, it then discards +the newly arrived message.</p> +<p>If you do not like throtteling, set the timeout to 0 - the message will then +immediately be discarded. If you use a high timeout, be sure you know what you +do. If a high main message queue enqueue timeout is set, it can lead to +something like a complete hang of the system. The same problem does not apply to +action queues.</p> +<h2>Rate Limiting</h2> +<p>Rate limiting provides a way to prevent rsyslogd from processing things too +fast. It can, for example, prevent overruning a receiver system.</p> +<p>Currently, there are only limited rate-limiting features available. The "<i>$<object>QueueDequeueSlowdown</i>" +directive allows to specify how long (in microseconds) dequeueing should be +delayed. While simple, it still is powerful. For example, using a +DequeueSlowdown delay of 1,000 microseconds on a UDP send action ensures that no +more than 1,000 messages can be sent within a second (actually less, as there is +also some time needed for the processing itself). </p> +<h2>Terminating Queues</h2> +<p>Terminating a process sounds easy, but can be complex. Terminting a running +queue is in fact the most complex operation a queue object can perform. You +don't see that from a user's point of view, but its quite hard work for the +developer to do everything in the right order.</p> +<p>The complexity arises when the queue has still data enqueued when it +finishes. Rsyslog tries to preserve as much of it as possible. As a first +measure, there is a regular queue time out ("<i>$<object>QueueTimeoutShutdown</i>", +specified in milliseconds): the queue workers are given that time period to +finish processing the queue.</p> +<p>If after that period there is still data in the queue, workers are instructed +to finish the current data element and then terminate. This essentially means +any other data is lost. There is another timeout ("<i>$<object>QueueTimeoutActionCompletion</i>", +also specified in milliseconds) that specifies how long the workers have to +finish the current element. If that timeout expires, any remaining workers are +cancelled and the queue is brought down.</p> +<p>If you do not like to lose data on shutdown, the "<i>$<object>QueueSaveOnShutdown</i>" +parameter can be set to "on". This requires either a disk or disk-assisted +queue. If set, rsyslogd ensures that any queue elements are saved to disk before +it terminates. This includes data elements there were begun being processed by +workers that needed to be cancelled due to too-long processing.</p> +<h1>Where are Queues Used?</h1> +<p> Currently, queues are used for the main message queue and for the +actions.</p> +<p>There is a single main message queue inside rsyslog. Each input module +delivers messages to it. The main message queue worker filters messages based on +rules specified in rsyslog.conf and dispatches them to the individual action +queues. Once a message is in an action queue, it is deleted from the main +message queue.</p> +<p>There are multiple action queues, one for each configured action. By default, +these queues operate in direct (non-queueing) mode. Action queues are fully +configurable and thus can be changed to whatever is best for the given use case.</p> +<p>Future versions of rsyslog will utilize queues at other places, too.</p> +<p>Not all queues necessarily support the full set of queue configuration +parameters, because not all are applicable. For example, in current output +module design, actions do not support multi-threading. Consequently, the number +of worker threads is fixed to one for action queues and can not be changed.</p> </body> |