NEWS: 2020-10-22: 0.6 is a new major version!
Sonnerie is a time-series database. Map a timestamp to a floating-point value. Store multiple of these series in a single database. Insert tens of millions of samples in minutes, on rotational media or solid-state.
Sonnerie is optimized for storing data that comes in as many values over many series, and for reading one series at a time. It is also very good at dumping lexicographically sequential series (which means: everything).
Sonnerie can very efficiently do random insertions and updates, and works well for huge databases. Due to the compact disk format, sparse data such as keys with only a few timestamps can be very efficiently stored.
- A straight-forward protocol for reading and writing
- Easy setup: insert data on the command line.
- No query language
- Transactional: a transaction is completely committed or not at all.
- Isolated: A transaction doesn't see updates from other transactions or expose its changes until it has been committed.
- Durable: committed data is resistant to loss from unexpected shutdown.
- Nanosecond-resolution timestamps (64 bit), 1970-2554
- No weird dependencies, no virtual machines, one single native binary for the command line tool
- floating point, integer, and string values, multiple columns per sample
Sonnerie runs on Unix-like systems and is developed on Linux.
Sonnerie is implemented in Rust, a systems programming language that runs
blazingly fast. Installation from source therefor requires you to
install the rust compiler,
which is as simple as:
curl https://sh.rustup.rs -sSf | sh.
Sonnerie can then be installed from Cargo:
cargo install sonnerie.
Sonnerie consists of one executable,
Create a database
Create a database by creating a directory and an empty file named "
mkdir database touch database/main
echo -e "\ fibonacci 2020-01-01T00:00:00 1 fibonacci 2020-01-02T00:00:00 1 fibonacci 2020-01-03T00:00:00 2 fibonacci 2020-01-04T00:00:00 3 fibonacci 2020-01-05T00:00:00 5 fibonacci 2020-01-06T00:00:00 8" \ | sonnerie -d database/ add --format u --timestamp-format=%FT%T
If the "add" command succeeds, then the transaction is committed to disk.
Items added with
sonnerie add must be sorted lexicographically by their
key and then chronologically. This requirement does not exist in
Read the data back
sonnerie -d database/ read %
% is a wildcard as is used in "
LIKE" in SQL and filters
on the key. Searching based on a prefix is very efficient:
sonnerie -d database/ read fib%
Sonnerie outputs the matched values:
fibonacci 2020-01-01 00:00:00 1 fibonacci 2020-01-02 00:00:00 1 fibonacci 2020-01-03 00:00:00 2 fibonacci 2020-01-04 00:00:00 3 fibonacci 2020-01-05 00:00:00 5 fibonacci 2020-01-06 00:00:00 8
Each series has a
format. The format is specified as a
bunch of single character codes, one for each value.
The character codes are:
f- a 32 bit float (f32)
F- a 64 bit float (f64)
u- a 32 bit unsigned integer (u32)
U- a 64 bit unsigned integer (u64)
i- a 32 bit signed integer (i32)
I- a 64 bit signed integer (i64)
s- a UTF-8 encoded string type. When strings are outputted, they are encoded in "backslash escaped" form, so all whitespace and backslashes are preceded by a backslash.
In the above "fibonacci" example, we're using the "u" format.
Multi-column rows are not extensively tested, but would look something like this, for two floating point values representing latitude and longitude:
oceanic-airlines 2018-01-01T00:00:00 ff 37.686751 -122.602227 oceanic-airlines 2018-01-01T00:00:01 ff 37.686810 -122.603713 oceanic-airlines 2018-01-01T00:00:02 ff 37.686873 -122.605997 oceanic-airlines 2018-01-01T00:00:03 ff 37.687022 -122.609997 oceanic-airlines 2018-01-01T00:00:04 ff 37.687364 -122.610945 oceanic-airlines 2018-01-01T00:00:05 ff 37.687503 -122.615211
Sonnerie allows heterogeneous formats.
A single key may change its format, for example:
keyname 2020-01-01T00:00:00 u 42 keyname 2020-01-02T00:00:00 f 3.1415 keyname 2020-01-03T00:00:00 s Now\ a\ string
While a key may change its format, it has more storage overhead, so it's best to not allow keys to oscillate between types.
This is permitted new in version 0.6, older versions had an "unsafe" mode that allowed the test to be bypassed for performance.
No server is necessary
All actions can be done by running
sonnerie -d /path/to/data/. Furthermore,
a file, (after it gets its ".tmp" suffix removed) will never change, though
the file named
main will get replaced sometimes. This means you can
replicate a database by hardlinking all the files (
The database must be compacted
On a regular (possibly daily) basis, you must compact the database. This rolls a bunch of transaction files into a single large transaction file. This is important for performance. By the time about 100 transaction files are present, performance suffers greatly. Therefor, compact the database at approximately the rate necessary to prevent that.
There are two types of compactions, a major and a minor one. A major
one replaces the entire database, which requires reading
and rewriting the entire database. A minor one replaces all of the transaction
files with a single new transaction file. This is a lot faster because it
requires only reading and rewriting the contents the transaction files
and not the
A major compaction is accomplished with:
sonnerie -d /path/to/data/ compact --major
And a minor compaction:
sonnerie -d /path/to/data/ compact
Compacting doesn't block readers or writers, but only one can happen at any given moment, so a lock is placed to prevent multiple concurrent compactions.
Compactions are atomic, so you can cancel it (with
^C) at any time.
You can compact and filter
In case some data in the database needs to be removed, you can use
compact with the
--gegnum option. Gegnum means "through" in Icelandic.
This command removes records that start with
compact --major --gegnum 'grep -v ^bad-objects'
Do a normal compaction, but also count records:
compact --major --gegnum 'pv -l'
--gegnum runs its command inside a /bin/sh, so pipelines work. Filter
out bad objects AND modify the names of other objects:
compact --major --gegnum 'grep -v ^bad-objects | sed "s/^old-name/new-name/"'
You can also see a preview of its output by piping your command into
| tee /dev/stderr.
Note that the rows come as "key\ttimestamp\tformat\tvalue"
You can also "read | filter | add" into a different database, but
you to modify an existing database which is useful for online maintenance on a database
that gets concurrent updates.
A server is provided so that you can conveniently read and write to the database via HTTP.
sonnerie-serve -d /path/to/database/ -l 0.0.0.0:5555 and then you may
Read the named series:
Read series by wildcard:
Output human-readable timestamps:
Add more data:
curl -X PUT http://localhost:5555/ --data-binary 'fibonacci 2020-01-07T00:00:00 u 13'
200 OK means that the transaction was committed)
sonnerie-serve allows unsorted input.
Note that because sonnerie
mmaps its files, sonnerie-serve will show
huge values for its virtual memory usage (
VIRT in top), but actual
memory utilization will be reasonable.
You may continue to read and modify your sonnerie database by the command
line or even via another concurrently-running
An alternate approach is to use "sshfs" to mount the database remotely. This approach is very performant because only compressed data goes through the network and the server doesn't need to do any of the decompressing. Avoid nfs because compactions will cause files to get deleted, and then the client will get an IO error, as NFS cannot track files that are closed on the server.
Sonnerie can be used as a Rust library so you can read and write databases directly, but the API is incomplete and poorly documented, for now.
Sonnerie is used in production
Sonnerie is used by e.ventures Management LLC with a >100GiB database and 10s of billions of rows.
An approximate average lookup time for a random key is around 100ms on an SSD and much slower on a busy rotational media device. Sequential access (i.e., reading the whole database in lexicographical order) is somewhere around 2k keys/sec and 1M records/sec, very much depending on the data itself.
Sonnerie was implemented by Charles Samuels at e.ventures Management LLC.