The Whisper Database

Whisper is a fixed-size database, similar in design and purpose to RRD (round-robin-database). It provides fast, reliable storage of numeric data over time. Whisper allows for higher resolution (seconds per point) of recent data to degrade into lower resolutions for long-term retention of historical data.

Data Points

Data points in Whisper are stored on-disk as big-endian double-precision floats. Each value is paired with a timestamp in seconds since the UNIX Epoch (01-01-1970). The data value is parsed by the Python float() function and as such behaves in the same way for special strings such as 'inf'. Maximum and minimum values are determined by the Python interpreter’s allowable range for float values which can be found by executing:

python -c 'import sys; print sys.float_info'

Archives: Retention and Precision

Whisper databases contain one or more archives, each with a specific data resolution and retention (defined in number of points or max timestamp age). Archives are ordered from the highest-resolution and shortest retention archive to the lowest-resolution and longest retention period archive.

To support accurate aggregation from higher to lower resolution archives, the precision of a longer retention archive must be divisible by precision of next lower retention archive. For example, an archive with 1 data point every 60 seconds can have a lower-resolution archive following it with a resolution of 1 data point every 300 seconds because 60 cleanly divides 300. In contrast, a 180 second precision (3 minutes) could not be followed by a 600 second precision (10 minutes) because the ratio of points to be propagated from the first archive to the next would be 3 1/3 and Whisper will not do partial point interpolation.

The total retention time of the database is determined by the archive with the highest retention as the time period covered by each archive is overlapping (see Multi-Archive Storage and Retrieval Behavior). That is, a pair of archives with retentions of 1 month and 1 year will not provide 13 months of data storage as may be guessed. Instead, it will provide 1 year of storage - the length of its longest archive.

Rollup Aggregation

Whisper databases with more than a single archive need a strategy to collapse multiple data points for when the data rolls up a lower precision archive. By default, an average function is used. Available aggregation methods are:

• average
• sum
• last
• max
• min

Multi-Archive Storage and Retrieval Behavior

When Whisper writes to a database with multiple archives, the incoming data point is written to all archives at once. The data point will be written to the highest resolution archive as-is, and will be aggregated by the configured aggregation method (see Rollup Aggregation) and placed into each of the higher-retention archives. If you are in need for aggregation of the highest resolution points, please consider using carbon-aggregator for that purpose.

When data is retrieved (scoped by a time range), the first archive which can satisfy the entire time period is used. If the time period overlaps an archive boundary, the lower-resolution archive will be used. This allows for a simpler behavior while retrieving data as the data’s resolution is consistent through an entire returned series.

Disk Space Efficiency

Whisper is somewhat inefficient in its usage of disk space because of certain design choices:

Each data point is stored with its timestamp

Rather than a timestamp being inferred from its position in the archive, timestamps are stored with each point. The timestamps are used during data retrieval to check the validity of the data point. If a timestamp does not match the expected value for its position relative to the beginning of the requested series, it is known to be out of date and a null value is returned

Archives overlap time periods

During the write of a data point, Whisper stores the same data in all archives at once (see Multi-Archive Storage and Retrieval Behavior). Implied by this behavior is that all archives store from now until each of their retention times. Because of this, lower-resolution archives should be configured to significantly lower resolution and higher retentions than their higher-resolution counterparts so as to reduce the overlap.

All time-slots within an archive take up space whether or not a value is stored

While Whisper allows for reliable storage of irregular updates, it is most space efficient when data points are stored at every update interval. This behavior is a consequence of the fixed-size design of the database and allows the reading and writing of series data to be performed in a single contiguous disk operation (for each archive in a database).

Differences Between Whisper and RRD

RRD can not take updates to a time-slot prior to its most recent update

This means that there is no way to back-fill data in an RRD series. Whisper does not have this limitation, and this makes importing historical data into Graphite much more simple and easy

RRD was not designed with irregular updates in mind

In many cases (depending on configuration) if an update is made to an RRD series but is not followed up by another update soon, the original update will be lost. This makes it less suitable for recording data such as operational metrics (e.g. code pushes)

Whisper requires that metric updates occur at the same interval as the finest resolution storage archive

This pushes the onus of aggregating values to fit into the finest precision archive to the user rather than the database. It also means that updates are written immediately into the finest precision archive rather than being staged first for aggregation and written later (during a subsequent write operation) as they are in RRD.

Performance

Whisper is fast enough for most purposes. It is slower than RRDtool primarily as a consequence of Whisper being written in Python, while RRDtool is written in C. The speed difference between the two in practice is quite small as much effort was spent to optimize Whisper to be as close to RRDtool’s speed as possible. Testing has shown that update operations take anywhere from 2 to 3 times as long as RRDtool, and fetch operations take anywhere from 2 to 5 times as long. In practice the actual difference is measured in hundreds of microseconds (10^-4) which means less than a millisecond difference for simple cases.

Database Format

WhisperFile	Header,Data
	Header	Metadata,ArchiveInfo+
		Metadata	aggregationType,maxRetention,xFilesFactor,archiveCount
		ArchiveInfo	Offset,SecondsPerPoint,Points
	Data	Archive+
		Archive	Point+
			Point	timestamp,value

Data types in Python’s struct format:

Metadata	!2LfL
ArchiveInfo	!3L
Point	!Ld