Understanding rsyslog Queues

Rsyslog uses queues whenever two activities need to be loosely coupled. With a queue, one part of the system “produces” something while another part “consumes” this something. The “something” is most often syslog messages, but queues may also be used for other purposes.

This document provides a good insight into technical details, operation modes and implications. In addition to it, an rsyslog queue concepts overview document exists which tries to explain queues with the help of some analogies. This may probably be a better place to start reading about queues. I assume that once you have understood that document, the material here will be much easier to grasp and look much more natural.

The most prominent example is the main message queue. Whenever rsyslog receives a message (e.g. locally, via UDP, TCP or in whatever else way), it places these messages into the main message queue. Later, it is dequeued by the rule processor, which then evaluates which actions are to be carried out. In front of each action, there is also a queue, which potentially de-couples the filter processing from the actual action (e.g. writing to file, database or forwarding to another host).

Where are Queues Used?

Currently, queues are used for the main message queue and for the actions.

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.

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.

Future versions of rsyslog will most probably utilize queues at other places, too.

Wherever “<object>”  is used in the config file statements, substitute “<object>” with either “MainMsg” or “Action”. The former will set main message queue parameters, the later parameters for the next action that will be created. Action queue parameters can not be modified once the action has been specified. For example, to tell the main message queue to save its content on shutdown, use $MainMsgQueueSaveOnShutdown on”.

If the same parameter is specified multiple times before a queue is created, the last one specified takes precedence. The main message queue is created after parsing the config file and all of its potential includes. An action queue is created each time an action selector is specified. Action queue parameters are reset to default after an action queue has been created (to provide a clean environment for the next action).

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.

Queue Modes

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. The queue mode (aka “type”) is set via the “$<object>QueueType” config directive.

Direct Queues

Direct queues are non-queuing queues. A queue in direct mode does neither queue nor buffer any of the queue elements but rather passes the element directly (and immediately) from the producer to the consumer. This sounds strange, but there is a good reason for this queue type.

Direct mode queues allow to use queues generically, even in places where queuing is not always desired. A good example is the queue in front of output actions. While it makes perfect sense to buffer forwarding actions or database writes, it makes only limited sense to build up a queue in front of simple local file writes. Yet, rsyslog still has a queue in front of every action. So for file writes, the queue mode can simply be set to “direct”, in which case no queuing happens.

Please note that a direct queue also is the only queue type that passes back the execution return code (success/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.

To create a direct queue, use the “$<object>QueueType Direct” config directive.

Disk Queues

Disk queues use disk drives for buffering. The important fact is that they always use the disk and do not buffer anything in memory. Thus, the queue is ultra-reliable, but by far the slowest mode. For regular use cases, this queue mode is not recommended. It is useful if log data is so important that it must not be lost, even in extreme cases.

When a disk queue is written, it is done in chunks. Each chunk receives its individual file. Files are named with a prefix (set via the “$<object>QueueFilename” config directive) and followed by a 7-digit number (starting at one and incremented for each file). Chunks are 10mb by default, a different size can be set via the”$<object>QueueMaxFileSize” config directive. Note that the size limit is not a sharp one: rsyslog always writes one complete queue entry, even if it violates the size limit. So chunks are actually a little bit (usually less than 1k) larger then the configured size. Each chunk also has a different size for the same reason. If you observe different chunk sizes, you can relax: this is not a problem.

Writing in chunks is used so that processed data can quickly be deleted and is free for other uses - while at the same time keeping no artificial upper limit on disk space used. If a disk quota is set (instructions further below), be sure that the quota/chunk size allows at least two chunks to be written. Rsyslog currently does not check that and will fail miserably if a single chunk is over the quota.

Creating new chunks costs performance but provides quicker ability to free disk space. The 10mb default is considered a good compromise between these two. However, it may make sense to adapt these settings to local policies. For example, if a disk queue is written on a dedicated 200gb disk, it may make sense to use a 2gb (or even larger) chunk size.

Please note, however, that the disk queue by default does not update its housekeeping structures every time it writes to disk. This is for performance reasons. In the event of failure, data will still be lost (except when manually 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.

Each queue can be placed on a different disk for best performance and/or isolation. This is currently selected by specifying different $WorkDirectory config directives before the queue creation statement.

To create a disk queue, use the “$<object>QueueType Disk” config directive. Checkpoint intervals can be specified via “$<object>QueueCheckpointInterval”, with 0 meaning no checkpoints. Note that disk-based queues can be made very reliable by issuing a (f)sync after each write operation. Starting with version 4.3.2, this can be requested via “<object>QueueSyncQueueFiles on/off with the default being off. Activating this option has a performance penalty, so it should not be turned on without reason.

If you happen to lose or otherwise need the housekeeping structures and have all yours queue chunks you can use perl script included in rsyslog package to generate it. Usage: recover_qi.pl -w $WorkDirectory -f QueueFileName -d 8 > QueueFileName.qi

In-Memory Queues

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. 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 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.

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.

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).

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.

In general, it is advised to use LinkedList mode if in doubt. The processing overhead compared to FixedArray is low and may be outweighed by the reduction in memory use. Paging in most-often-unused pointer array pages can be much slower than dynamically allocating them.

To create an in-memory queue, use the “$<object>QueueType LinkedList” or  “$<object>QueueType FixedArray” config directive.

Disk-Assisted Memory Queues

If a disk queue name is defined for in-memory queues (via $<object>QueueFileName), they automatically become “disk-assisted” (DA). In that mode, data is written to disk (and read back) on an as-needed basis.

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 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 disk space only) and data can be persisted between rsyslogd runs.

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].

DA queues are typically used to de-couple potentially long-running and unreliable actions (to make them reliable). For example, it is recommended 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 massive database inserts which nicely describes this use case. It may even be a good read if you do not intend to use databases.

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!).

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.

The water marks can be set via the “$<object>QueueHighWatermark” and  “$<object>QueueLowWatermark” configuration file directives. Note that these are actual numbers, not percentages. Be sure they make sense (also in respect to “$<object>QueueSize”). Rsyslodg does perform some checks on the numbers provided, and issues warning when numbers are “suspicious”.

Limiting the Queue Size

All queues, including disk queues, have a limit of the number of elements they can enqueue. This is set via the “$<object>QueueSize” 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).

Disk assisted queues are special in that they do not 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 “$<object>QueueMaxDiskSpace” 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).

In general, it is a good idea to limit the physical 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).

Worker Thread Pools

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).

Worker threads are started and stopped on an as-needed basis. On a system without activity, there may be no worker at all running. One is automatically started when a message comes in. Similarly, additional workers are started if the queue grows above a specific size. The “$<object>QueueWorkerThreadMinimumMessages”  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. Similarly, a third worker will be started when there are at least 300 messages, a forth when reaching 400 and so on.

It, however, does not make sense to have too many worker threads running in parallel. Thus, the upper limit can be set via “$<object>QueueWorkerThreads”. If it, for example, is set to four, no more than four workers will ever be started, no matter how many elements are enqueued.

Worker threads that have been started are kept running until an inactivity timeout happens. The timeout can be set via “$<object>QueueWorkerTimeoutThreadShutdown” 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. If you would like to never shutdown any worker threads, specify -1 for this parameter.

Discarding Messages

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 lose. 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.

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 consist of high-priority messages…

The discard watermark is set via the “$<object>QueueDiscardMark” directive. The priority of messages to be discarded is set via “$<object>QueueDiscardSeverity”. This directive accepts both the usual textual severity as well as a numerical one. To understand it, you must be aware of the numerical severity values. They are defined in RFC 3164:

Code Severity
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

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. An alternative is to specify the numerical value 8 as DiscardSeverity. This is also the default setting to prevent unintentional message loss. So if you would like to use message discarding, you need to set” $<object>QueueDiscardSeverity” to an actual value.

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 absolutely unavoidable and you prefer to discard less important messages first.

Filled-Up Queues

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 resolution for this scenario.

During throttling, 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 throttling is highly undesirable for unreliable sources, like UDP message reception. So it is not a good thing to run into throttling mode at all.

We can not hold processing infinitely, not even when throttling. For example, throttling 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 (“$<object>QueueTimeoutEnqueue”, specified in milliseconds) if no space becomes available. As a last resort, it then discards the newly arrived message.

If you do not like throttling, 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.

Rate Limiting

Rate limiting provides a way to prevent rsyslogd from processing things too fast. It can, for example, prevent overrunning a receiver system.

Currently, there are only limited rate-limiting features available. The “$<object>QueueDequeueSlowdown”  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).

Processing Timeframes

Queues can be set to dequeue (process) messages only during certain timeframes. This is useful if you, for example, would like to transfer the bulk of messages only during off-peak hours, e.g. when you have only limited bandwidth on the network path to the central server.

Currently, only a single timeframe is supported and, even worse, it can only be specified by the hour. It is not hard to extend rsyslog’s capabilities in this regard - it was just not requested so far. So if you need more fine-grained control, let us know and we’ll probably implement it. There are two configuration directives, both should be used together or results are unpredictable:” $<object>QueueDequeueTimeBegin <hour>” and ”$<object>QueueDequeueTimeEnd <hour>”. The hour parameter must be specified in 24-hour format (so 10pm is 22). A use case for this parameter can be found in the rsyslog wiki.

Performance

The locking involved with maintaining the queue has a potentially large performance impact. How large this is, and if it exists at all, depends much on the configuration and actual use case. However, the queue is able to work on so-called “batches” when dequeueing data elements. With batches, multiple data elements are dequeued at once (with a single locking call). The queue dequeues all available elements up to a configured upper limit (<object>DequeueBatchSize <number>). It is important to note that the actual upper limit is dictated by availability. The queue engine will never wait for a batch to fill. So even if a high upper limit is configured, batches may consist of fewer elements, even just one, if there are no more elements waiting in the queue.

Batching can improve performance considerably. Note, however, that it affects the order in which messages are passed to the queue worker threads, as each worker now receive as batch of messages. Also, the larger the batch size and the higher the maximum number of permitted worker threads, the more main memory is needed. For a busy server, large batch sizes (around 1,000 or even more elements) may be useful. Please note that with batching, the main memory must hold BatchSize * NumOfWorkers objects in memory (worst-case scenario), even if running in disk-only mode. So if you use the default 5 workers at the main message queue and set the batch size to 1,000, you need to be prepared that the main message queue holds up to 5,000 messages in main memory in addition to the configured queue size limits!

The queue object’s default maximum batch size is eight, but there exists different defaults for the actual parts of rsyslog processing that utilize queues. So you need to check these object’s defaults.

Terminating Queues

Terminating a process sounds easy, but can be complex. Terminating 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.

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 (“$<object>QueueTimeoutShutdown”, specified in milliseconds): the queue workers are given that time period to finish processing the queue.

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 (“$<object>QueueTimeoutActionCompletion”, 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.

If you do not like to lose data on shutdown, the “$<object>QueueSaveOnShutdown” 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. For a large queue, this operation may be lengthy. No timeout applies to a required shutdown save.