This repo contains code for benchmarking several time series databases, including TimescaleDB, MongoDB, InfluxDB, CrateDB and Cassandra. This code is based on a fork of work initially made public by InfluxDB at https://github.com/influxdata/influxdb-comparisons.
Current databases supported:
- Akumuli (supplemental docs)
- Cassandra (supplemental docs)
- ClickHouse (supplemental docs)
- CrateDB (supplemental docs)
- InfluxDB (supplemental docs)
- MongoDB (supplemental docs)
- QuestDB (supplemental docs)
- SiriDB (supplemental docs)
- TimescaleDB (supplemental docs)
- Timestream (supplemental docs)
- VictoriaMetrics (supplemental docs)
- TDengine (supplemental docs)
The Time Series Benchmark Suite (TSBS) is a collection of Go programs that are used to generate datasets and then benchmark read and write performance of various databases. The intent is to make the TSBS extensible so that a variety of use cases (e.g., devops, IoT, finance, etc.), query types, and databases can be included and benchmarked. To this end we hope to help prospective database administrators find the best database for their needs and their workloads. Further, if you are the developer of a time series database and want to include your database in the TSBS, feel free to open a pull request to add it!
Currently, TSBS supports two use cases.
A 'dev ops' use case, which comes in two forms. The full form is used to generate, insert, and measure data from 9 'systems' that could be monitored in a real world dev ops scenario (e.g., CPU, memory, disk, etc). Together, these 9 systems generate 100 metrics per reading interval. The alternate form focuses solely on CPU metrics for a simpler, more streamlined use case. This use case generates 10 CPU metrics per reading.
In addition to metric readings, 'tags' (including the location
of the host, its operating system, etc) are generated for each host
with readings in the dataset. Each unique set of tags identifies
one host in the dataset and the number of different hosts generated is
defined by the scale flag (see below).
The second use case is meant to simulate the data load in an IoT environment. This use case simulates data streaming from a set of trucks belonging to a fictional trucking company. This use case simulates diagnostic data and metrics from each truck, and introduces environmental factors such as out-of-order data and batch ingestion (for trucks that are offline for a period of time). It also tracks truck metadata and uses this to tie metrics and diagnostics together as part of the query set.
The queries that are generated as part of this use case will cover both real time truck status and analytics that will look at the time series data in an effort to be more predictive about truck behavior. The scale factor with this use case will be based on the number of trucks tracked.
Not all databases implement all use cases. This table below shows which use cases are implemented for each database:
| Database | Dev ops | IoT |
|---|---|---|
| Akumuli | X¹ | |
| Cassandra | X | |
| ClickHouse | X | |
| CrateDB | X | |
| InfluxDB | X | X |
| MongoDB | X | |
| QuestDB | X | X |
| SiriDB | X | |
| TimescaleDB | X | X |
| Timestream | X | |
| VictoriaMetrics | X² | |
| TDengine | X | X |
¹ Does not support the groupby-orderby-limit query
² Does not support the groupby-orderby-limit, lastpoint, high-cpu-1, high-cpu-all queries
TSBS is used to benchmark bulk load performance and
query execution performance. (It currently does not measure
concurrent insert and query performance, which is a future priority.)
To accomplish this in a fair way, the data to be inserted and the
queries to run are pre-generated and native Go clients are used
wherever possible to connect to each database (e.g., mgo for MongoDB,
aws sdk for Timestream).
Although the data is randomly generated, TSBS data and queries are entirely deterministic. By supplying the same PRNG (pseudo-random number generator) seed to the generation programs, each database is loaded with identical data and queried using identical queries.
TSBS is a collection of Go programs (with some auxiliary bash and Python scripts).
Notice
TDengine uses the go connector to perform data writing and querying in the TSBS test, and the installation process relies on the C client library. Therefore, prior to compiling TSBS, the taos client must be installed to obtain the necessary library files. For instructions on how to install the taos client and server, refer to the TDengine documentation .
The easiest way to get and install the Go programs is to use
go get and then make all to install all binaries:
# Fetch TSBS and its dependencies
$ go get github.com/taosdata/tsbs
$ cd $GOPATH/src/github.com/taosdata/tsbs
$ makeUsing TSBS for benchmarking involves 3 phases: data and query generation, data loading/insertion, and query execution.
So that benchmarking results are not affected by generating data or queries on-the-fly, with TSBS you generate the data and queries you want to benchmark first, and then you can (re-)use it as input to the benchmarking phases.
Variables needed:
- a use case. E.g.,
iot(choose fromcpu-only,devops, oriot) - a PRNG seed for deterministic generation. E.g.,
123 - the number of devices / trucks to generate for. E.g.,
4000 - a start time for the data's timestamps. E.g.,
2016-01-01T00:00:00Z - an end time. E.g.,
2016-01-04T00:00:00Z - how much time should be between each reading per device, in seconds. E.g.,
10s - and which database(s) you want to generate for. E.g.,
timescaledb(choose fromcassandra,clickhouse,cratedb,influx,mongo,questdb,siridb,timescaledborvictoriametrics)
Given the above steps you can now generate a dataset (or multiple
datasets, if you chose to generate for multiple databases) that can
be used to benchmark data loading of the database(s) chosen using
the tsbs_generate_data tool:
$ tsbs_generate_data --use-case="iot" --seed=123 --scale=4000 \
--timestamp-start="2016-01-01T00:00:00Z" \
--timestamp-end="2016-01-04T00:00:00Z" \
--log-interval="10s" --format="timescaledb" \
| gzip > /tmp/timescaledb-data.gz
# Each additional database would be a separate call.Note: We pipe the output to gzip to reduce on-disk space. This also requires you to pipe through gunzip when you run your tests.
The example above will generate a pseudo-CSV file that can be used to bulk load data into TimescaleDB. Each database has it's own format of how it stores the data to make it easiest for its corresponding loader to write data. The above configuration will generate just over 100M rows (1B metrics), which is usually a good starting point. Increasing the time period by a day will add an additional ~33M rows so that, e.g., 30 days would yield a billion rows (10B metrics)
The main difference between the iot use case and other use cases is that
it generates data which can contain out-of-order, missing, or empty
entries to better represent real-life scenarios associated to the use case.
Using a specified seed means that we can do this in a deterministic and
reproducible way for multiple runs of data generation.
Variables needed:
- the same use case, seed, # of devices, and start time as used in data generation
- an end time that is one second after the end time from data generation. E.g., for
2016-01-04T00:00:00Zuse2016-01-04T00:00:01Z - the number of queries to generate. E.g.,
1000 - and the type of query you'd like to generate. E.g.,
single-groupby-1-1-1orlast-loc
For the last step there are numerous queries to choose from, which are
listed in Appendix I. Additionally, the file
scripts/generate_queries.sh contains a list of all of them as the
default value for the environmental variable QUERY_TYPES. If you are
generating more than one type of query, we recommend you use the
helper script.
For generating just one set of queries for a given type:
$ tsbs_generate_queries --use-case="iot" --seed=123 --scale=4000 \
--timestamp-start="2016-01-01T00:00:00Z" \
--timestamp-end="2016-01-04T00:00:01Z" \
--queries=1000 --query-type="breakdown-frequency" --format="timescaledb" \
| gzip > /tmp/timescaledb-queries-breakdown-frequency.gzNote: We pipe the output to gzip to reduce on-disk space. This also requires you to pipe through gunzip when you run your tests.
For generating sets of queries for multiple types:
$ FORMATS="timescaledb" SCALE=4000 SEED=123 \
TS_START="2016-01-01T00:00:00Z" \
TS_END="2016-01-04T00:00:01Z" \
QUERIES=1000 QUERY_TYPES="last-loc low-fuel avg-load" \
BULK_DATA_DIR="/tmp/bulk_queries" scripts/generate_queries.shA full list of query types can be found in Appendix I at the end of this README.
TSBS has two ways to benchmark insert/write performance:
- On the fly simulation and load with
tsbs_load - Pre-generate data to a file and load it either with
tsbs_loador the db specific executablestsbs_load_*
The tsbs_load executable can load data in any of the supported databases.
It can use a pregenerated data file as input, or simulate the data on the
fly.
You first start by generating a config yaml file populated with the default values for each property with:
$ tsbs_load config --target=<db-name> --data-source=[FILE|SIMULATOR]for example, to generate an example for TimescaleDB, loading the data from file
$ tsbs_load config --target=timescaledb --data-source=FILE
Wrote example config to: ./config.yamlYou can then run tsbs_load with the generated config file with:
$ tsbs_load load timescaledb --config=./config.yamlFor more details on how to use tsbs_load check out the supplemental docs
TSBS measures insert/write performance by taking the data generated in
the previous step and using it as input to a database-specific command
line program. To the extent that insert programs can be shared, we have
made an effort to do that (e.g., the TimescaleDB loader can
be used with a regular PostgreSQL database if desired). Each loader does
share some common flags -- e.g., batch size (number of readings inserted
together), workers (number of concurrently inserting clients), connection
details (host & ports), etc -- but they also have database-specific tuning
flags. To find the flags for a particular database, use the -help flag
(e.g., tsbs_load_timescaledb -help).
Here's an example of loading data to a remote timescaledb instance with SSL required, with a gzipped data set as created in the instructions above:
cat /tmp/timescaledb-data.gz | gunzip | tsbs_load_timescaledb \
--postgres="sslmode=require" --host="my.tsdb.host" --port=5432 --pass="password" \
--user="benchmarkuser" --admin-db-name=defaultdb --workers=8 \
--in-table-partition-tag=true --chunk-time=8h --write-profile= \
--field-index-count=1 --do-create-db=true --force-text-format=false \
--do-abort-on-exist=falseFor simpler testing, especially locally, we also supply
scripts/load/load_<database>.sh for convenience with many of the flags set
to a reasonable default for some of the databases.
So for loading into TimescaleDB, ensure that TimescaleDB is running and
then use:
# Will insert using 2 clients, batch sizes of 10k, from a file
# named `timescaledb-data.gz` in directory `/tmp`
$ NUM_WORKERS=2 BATCH_SIZE=10000 BULK_DATA_DIR=/tmp \
scripts/load/load_timescaledb.shThis will create a new database called benchmark where the data is
stored. It will overwrite the database if it exists; if you don't
want that to happen, supply a different DATABASE_NAME to the above
command.
Example for writing to remote host using load_timescaledb.sh:
# Will insert using 2 clients, batch sizes of 10k, from a file
# named `timescaledb-data.gz` in directory `/tmp`
$ NUM_WORKERS=2 BATCH_SIZE=10000 BULK_DATA_DIR=/tmp DATABASE_HOST=remotehostname
DATABASE_USER=user DATABASE \
scripts/load/load_timescaledb.shBy default, statistics about the load performance are printed every 10s, and when the full dataset is loaded the looks like this:
time,per. metric/s,metric total,overall metric/s,per. row/s,row total,overall row/s
# ...
1518741528,914996.143291,9.652000E+08,1096817.886674,91499.614329,9.652000E+07,109681.788667
1518741548,1345006.018902,9.921000E+08,1102333.152918,134500.601890,9.921000E+07,110233.315292
1518741568,1149999.844750,1.015100E+09,1103369.385320,114999.984475,1.015100E+08,110336.938532
Summary:
loaded 1036800000 metrics in 936.525765sec with 8 workers (mean rate 1107070.449780/sec)
loaded 103680000 rows in 936.525765sec with 8 workers (mean rate 110707.044978/sec)
All but the last two lines contain the data in CSV format, with column names in the header. Those column names correspond to:
- timestamp,
- metrics per second in the period,
- total metrics inserted,
- overall metrics per second,
- rows per second in the period,
- total number of rows,
- overall rows per second.
For databases, like Cassandra, that do not use rows when inserting,
the last three values are always empty (indicated with a -).
The last two lines are a summary of how many metrics (and rows where applicable) were inserted, the wall time it took, and the average rate of insertion.
To measure query execution performance in TSBS, you first need to load
the data using the previous section and generate the queries as
described earlier. Once the data is loaded and the queries are generated,
just use the corresponding tsbs_run_queries_ binary for the database
being tested:
$ cat /tmp/queries/timescaledb-cpu-max-all-eight-hosts-queries.gz | \
gunzip | tsbs_run_queries_timescaledb --workers=8 \
--postgres="host=localhost user=postgres sslmode=disable"You can change the value of the --workers flag to
control the level of parallel queries run at the same time. The
resulting output will look similar to this:
run complete after 1000 queries with 8 workers:
TimescaleDB max cpu all fields, rand 8 hosts, rand 12hr by 1h:
min: 51.97ms, med: 757.55, mean: 2527.98ms, max: 28188.20ms, stddev: 2843.35ms, sum: 5056.0sec, count: 2000
all queries :
min: 51.97ms, med: 757.55, mean: 2527.98ms, max: 28188.20ms, stddev: 2843.35ms, sum: 5056.0sec, count: 2000
wall clock time: 633.936415sec
The output gives you the description of the query and multiple groupings of measurements (which may vary depending on the database).
For easier testing of multiple queries, we provide
scripts/generate_run_script.py which creates a bash script with commands
to run multiple query types in a row. The queries it generates should be
put in a file with one query per line and the path given to the script.
For example, if you had a file named queries.txt that looked like this:
last-loc
avg-load
high-load
long-driving-session
You could generate a run script named query_test.sh:
# Generate run script for TimescaleDB, using queries in `queries.txt`
# with the generated query files in /tmp/queries for 8 workers
$ python generate_run_script.py -d timescaledb -o /tmp/queries \
-w 8 -f queries.txt > query_test.shAnd the resulting script file would look like:
#!/bin/bash
# Queries
cat /tmp/queries/timescaledb-last-loc-queries.gz | gunzip | query_benchmarker_timescaledb --workers=8 --limit=1000 --hosts="localhost" --postgres="user=postgres sslmode=disable" | tee query_timescaledb_timescaledb-last-loc-queries.out
cat /tmp/queries/timescaledb-avg-load-queries.gz | gunzip | query_benchmarker_timescaledb --workers=8 --limit=1000 --hosts="localhost" --postgres="user=postgres sslmode=disable" | tee query_timescaledb_timescaledb-avg-load-queries.out
cat /tmp/queries/timescaledb-high-load-queries.gz | gunzip | query_benchmarker_timescaledb --workers=8 --limit=1000 --hosts="localhost" --postgres="user=postgres sslmode=disable" | tee query_timescaledb_timescaledb-high-load-queries.out
cat /tmp/queries/timescaledb-long-driving-session-queries.gz | gunzip | query_benchmarker_timescaledb --workers=8 --limit=1000 --hosts="localhost" --postgres="user=postgres sslmode=disable" | tee query_timescaledb_timescaledb-long-driving-session-queries.outAdditionally each tsbs_run_queries_ binary allows you print the
actual query results so that you can compare across databases that the
results are the same. Using the flag -print-responses will return
the results.
| Query type | Description |
|---|---|
| single-groupby-1-1-1 | Simple aggregrate (MAX) on one metric for 1 host, every 5 mins for 1 hour |
| single-groupby-1-1-12 | Simple aggregrate (MAX) on one metric for 1 host, every 5 mins for 12 hours |
| single-groupby-1-8-1 | Simple aggregrate (MAX) on one metric for 8 hosts, every 5 mins for 1 hour |
| single-groupby-5-1-1 | Simple aggregrate (MAX) on 5 metrics for 1 host, every 5 mins for 1 hour |
| single-groupby-5-1-12 | Simple aggregrate (MAX) on 5 metrics for 1 host, every 5 mins for 12 hours |
| single-groupby-5-8-1 | Simple aggregrate (MAX) on 5 metrics for 8 hosts, every 5 mins for 1 hour |
| cpu-max-all-1 | Aggregate across all CPU metrics per hour over 1 hour for a single host |
| cpu-max-all-8 | Aggregate across all CPU metrics per hour over 1 hour for eight hosts |
| double-groupby-1 | Aggregate on across both time and host, giving the average of 1 CPU metric per host per hour for 24 hours |
| double-groupby-5 | Aggregate on across both time and host, giving the average of 5 CPU metrics per host per hour for 24 hours |
| double-groupby-all | Aggregate on across both time and host, giving the average of all (10) CPU metrics per host per hour for 24 hours |
| high-cpu-all | All the readings where one metric is above a threshold across all hosts |
| high-cpu-1 | All the readings where one metric is above a threshold for a particular host |
| lastpoint | The last reading for each host |
| groupby-orderby-limit | The last 5 aggregate readings (across time) before a randomly chosen endpoint |
| Query type | Description |
|---|---|
| last-loc | Fetch real-time (i.e. last) location of each truck |
| low-fuel | Fetch all trucks with low fuel (less than 10%) |
| high-load | Fetch trucks with high current load (over 90% load capacity) |
| stationary-trucks | Fetch all trucks that are stationary (low avg velocity in last 10 mins) |
| long-driving-sessions | Get trucks which haven't rested for at least 20 mins in the last 4 hours |
| long-daily-sessions | Get trucks which drove more than 10 hours in the last 24 hours |
| avg-vs-projected-fuel-consumption | Calculate average vs. projected fuel consumption per fleet |
| avg-daily-driving-duration | Calculate average daily driving duration per driver |
| avg-daily-driving-session | Calculate average daily driving session per driver |
| avg-load | Calculate average load per truck model per fleet |
| daily-activity | Get the number of hours truck has been active (vs. out-of-commission) per day per fleet |
| breakdown-frequency | Calculate breakdown frequency by truck model |
We welcome contributions from the community to make TSBS better!
You can help either by opening an issue with any suggestions or bug reports, or by forking this repository, making your own contribution, and submitting a pull request.
Before we accept any contributions, Timescale contributors need to sign the Contributor License Agreement (CLA). By signing a CLA, we can ensure that the community is free and confident in its ability to use your contributions.