Under certain scenarios it is possible for logs or data to be ingested or created faster than the ability to flush it to some destinations. One such common scenario is when reading from big log files, especially with a large backlog, and dispatching the logs to a backend over the network, which takes time to respond. This generates backpressure leading to high memory consumption in the service.
In order to avoid backpressure, Fluent Bit implements a mechanism in the engine that restricts the amount of data that an input plugin can ingest, this is done through the configuration parameter Mem_Buf_Limit.
As described in the Buffering concepts section, Fluent Bit offers a hybrid mode for data handling: in-memory and filesystem (optional).
In memory is always available and can be restricted with Mem_Buf_Limit. If memory reaches this limit and you reach a backpressure scenario, you will not be able to ingest more data until the data chunks that are in memory can be flushed.
Depending on the input plugin in use, this might lead to discard incoming data (e.g: TCP input plugin). This can be mitigated by configuring secondary storage on the filesystem using the storage.type of filesystem (as described in Buffering & Storage). When the limit is reached, all the new data will be stored safely in the filesystem.
This option is disabled by default and can be applied to all input plugins. Let's explain its behavior using the following scenario:
Mem_Buf_Limit is set to 1MB (one megabyte)
input plugin tries to append 700KB
engine route the data to an output plugin
output plugin backend (HTTP Server) is down
engine scheduler will retry the flush after 10 seconds
input plugin tries to append 500KB
At this exact point, the engine will allow appending those 500KB of data into the memory; in total it will have 1.2MB of data buffered. The limit is permissive and will allow a single write past the limit, but once the limit is exceeded the following actions are taken:
block local buffers for the input plugin (cannot append more data)
notify the input plugin invoking a pause callback
The engine will protect itself and will not append more data coming from the input plugin in question; note that it is the responsibility of the plugin to keep state and decide what to do in that paused state.
After some time, usually measured in seconds, if the scheduler was able to flush the initial 700KB of data or it has given up after retrying, that amount of memory is released and the following actions will occur:
Upon data buffer release (700KB), the internal counters get updated
Counters now are set at 500KB
Since 500KB is < 1MB it checks the input plugin state
If the plugin is paused, it invokes a resume callback
input plugin can continue appending more data
Each plugin is independent and not all of them implements the pause and resume callbacks. As said, these callbacks are just a notification mechanism for the plugin.
One example of a plugin that implements these callbacks and keeps state correctly is the Tail Input plugin. When the pause callback is triggered, it pauses its collectors and stops appending data. Upon resume, it resumes the collectors and continues ingesting data.