Buffer plugins are used by output plugins. For example,
buf_file plugin by default to store incoming stream temporally before transmitting to S3.
Buffer plugins are, as you can tell by the name, pluggable. So you can choose a suitable backend based on your system requirements.
A buffer is essentially a set of "chunks". A chunk is a collection of records concatenated into a single blob. Chunks are periodically flushed to the output queue and then sent to the specified destination. Here is the diagram of how it works:
Buffer plugins allow fine-grained controls over the buffering behaviours through config options.
This option specifies which plugin to use as the backend.
In default installations, supported values are
The maximum size of a chunk allowed (default: 8MB)
If a chunk grows more than the limit, it gets flushed to the output
The maximum length of the output queue (default: 256)
If the limit gets exceeded, Fluentd will invoke an error handling
mechanism (See "Handling queue overflow" below for details).
The interval in seconds to wait before invoking the next buffer
flush (default: 60)
buffer_queue_full_action option controls the behaviour when the queue becomes full. For now, three modes are supported:
This mode raises a
BufferQueueLimitError exception to the
This mode is suitable for data streaming.
It's up to the input plugin to decide how to handle raised
This mode blocks the thread until the free space is vacated.
This mode is good for batch-like use-case.
DO NOT use this option casually just for avoiding
BufferQueueLimitError exceptions (see the deployment tips
This mode drops the oldest chunks.
This mode might be useful for monitoring systems, since newer
events are much more important than the older ones in this
If your Fluentd daemon experiences overflows frequently, it means that your destination is insufficient for your traffic. There are several measures you can take:
Upgrade the destination node to provide enough data-processing
@ERROR label to route overflowed events to another backup
<secondary> tag to route overflowed events to another backup
The chunks in the output queue are written out to the destination one by one. However, the problem is that there might occur an error while writing out a chunk. For such cases, buffer plugins are equipped with a "retry" mechanism that handles write failures gracefully.
Here is how it works. If Fluentd fails to write out a chunk, the chunk will not be purged from the queue, and then, after a certain interval, Fluentd will retry to write the chunk again. The intervals between retry attempts are determined by the exponential backoff algorithm, and we can control the behaviour finely through the following options:
The maximum number of retries for sending a chunk (default: 17)
retry_limit is exceeded, Fluentd will discard the given chunk.
If set true, it disables
retry_limit and make Fluentd retry
indefinitely (default: false).
The number of seconds the first retry will wait (default: 1.0)
This is the seed value used by the exponential backoff algorithm;
The wait interval will be doubled on each retry.
The maximum interval seconds to wait between retries (default:
If the wait interval reaches this limit, the exponentiation stops.
Buffer plugins support a special mode that groups the incoming data by time frames. For example, you can group the incoming access logs by date and save them to separate files. Under this mode, a buffer plugin will behave quite differently in a few key aspects:
It supports the additional
time_slice_* options (See Q1/Q2 below
Chunks will be flushed "lazily" based on the settings of the
time_slice_format option. See Q2 for the relationship of this
The default value of
buffer_chunk_limit becomes 256mb.
Normally, the output plugin determines in which mode the buffer plugin operates. For example,
out_file will enable the time-slicing mode. For the list of output plugins which enable the time-slicing mode, see this page.
This is done through the
time_slice_format option, which is set to "%Y%m%d" (daily) by default. If you want your chunks to be hourly, "%Y%m%d%H" will do the job.
For example, what happens to an event, timestamped at 2013-01-01 02:59:45 UTC, comes in at 2013-01-01 03:00:15 UTC? Would it make into the 2013-01-01 02:00:00-02:59:59 chunk?
This issue is addressed by setting the
time_slice_wait sets, in seconds, how long fluentd waits to accept "late" events into the chunk past the max time corresponding to that chunk. The default value is 600, which means it waits for 10 minutes before moving on. So, in the current example, as long as the events come in before 2013-01-01 03:10:00, it will make it in to the 2013-01-01 02:00:00-02:59:59 chunk.
Alternatively, you can also flush the chunks regularly using
flush_interval. Note that
time_slice_wait are mutually exclusive. If you set
time_slice_wait will be ignored and fluentd would issue a warning.
If this article is incorrect or outdated, or omits critical information, please let us know. Fluentd is a open source project under Cloud Native Computing Foundation (CNCF). All components are available under the Apache 2 License.