MarketStore

Read this in 日本語(Japanese)

Introduction

MarketStore is a database server optimized for financial time-series data. You can think of it as an extensible DataFrame service that is accessible from anywhere in your system, at higher scalability.

It is designed from the ground up to address scalability issues around handling large amounts of financial market data used in algorithmic trading backtesting, charting, and analyzing price history with data spanning many years, and granularity down to tick-level for the all US equities or the exploding cryptocurrencies space. If you are struggling with managing lots of HDF5 files, this is perfect solution to your problem.

The batteries are included with the basic install - you can start pulling crypto price data from GDAX and writing it to the db with a simple plugin configuration.

MarketStore enables you to query DataFrame content over the network at as low latency as your local HDF5 files from disk, and appending new data to the end is two orders of magnitude faster than DataFrame would be. This is because the storage format is optimized for the type of data and use cases as well as for modern filesystem/hardware characteristics.

MarketStore is production ready! At Alpaca it has been used in production for years in serious business. If you encounter a bug or are interested in getting involved, please see the contribution section for more details.

Install

Docker

If you want to get started right away, you can bootstrap a marketstore db instance using our latest docker image. The image comes pre-loaded with the default mkts.yml file and declares the VOLUME /data, as its root directory. To run the container with the defaults:

docker run -i -p 5993:5993 alpacamarkets/marketstore:latest

If you want to run a custom mkts.yml you can create a new container and load your mkts.yml file into it:

docker create --name mktsdb -p 5993:5993 alpacamarkets/marketstore:latest
docker cp mkts.yml mktsdb:/etc/mkts.yml
docker start -i mktsdb

You can also bind mount the container to a local host config file: a custom mkts.yml:

docker run -v /full/path/to/mkts.yml:/etc/mkts.yml -i -p 5993:5993 alpacamarkets/marketstore:latest

This allows you to test out the image included plugins with ease if you prefer to skip the copying step suggested above.

By default the container will not persist any written data to your container's host storage. To accomplish this, bind the data directory to a local location:

docker run -v "/path/to/store/data:/data" -i -p 5993:5993 alpacamarkets/marketstore:latest

Once data is written to the server you should see a file tree layout like the following that will persist across container runs:

>>> tree /<path_to_data>/marketstore
/<path_to_data>/marketstore
├── category_name
├── WALFile.1590868038674814776.walfile
├── SYMBOL_1
├── SYMBOL_2
├── SYMBOL_3

If you have built the cmd package locally, you can open a session with your running docker instance using:

marketstore connect --url localhost:5993

Source

MarketStore is implemented in Go, so you can install it by go install.

go install github.com/alpacahq/marketstore/v4@latest
# export GOROOT=$HOME/go
# export PATH=$PATH:$GOROOT/bin
marketstore --version

You can build it from source easily too. You need Go 1.11+ as it uses go mod to manage dependencies.

go get -u github.com/alpacahq/marketstore

then compile and install the project binaries using

make install

Optionally, you can install the project's included plugins using

make plugins

Homebrew on macOS

You can also install marketstore using the Homebrew package manager for macOS.

$ brew tap zjhmale/marketstore
$ brew install --HEAD marketstore

To upgrade marketstore in the future, use upgrade instead of install.

Then you are equipped with marketstore service plist also

$ brew services start marketstore
$ brew services stop marketstore

Usage

You can list available commands by running

marketstore

or

$GOPATH/bin/marketstore

depending on your GOPATH.

You can create a new configuration file named mkts.yml, populated with defaults by running:

$GOPATH/bin/marketstore init

and then start the marketstore server with:

$GOPATH/bin/marketstore start

The output will look something like:

example@alpaca:~/go/bin/src/github.com/alpacahq/marketstore$ marketstore
I0619 16:29:30.102101    7835 log.go:14] Disabling "enable_last_known" feature until it is fixed...
I0619 16:29:30.102980    7835 log.go:14] Initializing MarketStore...
I0619 16:29:30.103092    7835 log.go:14] WAL Setup: initCatalog true, initWALCache true, backgroundSync true, WALBypass false:
I0619 16:29:30.103179    7835 log.go:14] Root Directory: /example/go/bin/src/github.com/alpacahq/marketstore/project/data/mktsdb
I0619 16:29:30.144461    7835 log.go:14] My WALFILE: WALFile.1529450970104303654.walfile
I0619 16:29:30.144486    7835 log.go:14] Found a WALFILE: WALFile.1529450306968096708.walfile, entering replay...
I0619 16:29:30.244778    7835 log.go:14] Beginning WAL Replay
I0619 16:29:30.244861    7835 log.go:14] Partial Read
I0619 16:29:30.244882    7835 log.go:14] Entering replay of TGData
I0619 16:29:30.244903    7835 log.go:14] Replay of WAL file /example/go/bin/src/github.com/alpacahq/marketstore/project/data/mktsdb/WALFile.1529450306968096708.walfile finished
I0619 16:29:30.289401    7835 log.go:14] Finished replay of TGData
I0619 16:29:30.340760    7835 log.go:14] Launching rpc data server...
I0619 16:29:30.340792    7835 log.go:14] Initializing websocket...
I0619 16:29:30.340814    7835 plugins.go:14] InitializeTriggers
I0619 16:29:30.340824    7835 plugins.go:42] InitializeBgWorkers

Configuration

In order to run MarketStore, a YAML config file is needed. A default file (mkts.yml) can be created using marketstore init. The path to this file is passed in to the start command with the --config flag, or by default it finds a file named mkts.yml in the directory it is running from.

Options

Default mkts.yml

root_directory: data
listen_port: 5993
grpc_listen_port: 5995
log_level: info
stop_grace_period: 0
wal_rotate_interval: 5
stale_threshold: 5

Clients

After starting up a MarketStore instance on your machine, you're all set to be able to read and write tick data.

Python

pymarketstore is the standard python client. Make sure that in another terminal, you have marketstore running

• query data

import pymarketstore as pymkts
param = pymkts.Params('BTC', '1Min', 'OHLCV', limit=10)
cli = pymkts.Client()
reply = cli.query(param)
reply.first().df()

shows

Out[5]:
                               Open      High       Low     Close     Volume
Epoch
2018-01-17 17:19:00+00:00  10400.00  10400.25  10315.00  10337.25   7.772154
2018-01-17 17:20:00+00:00  10328.22  10359.00  10328.22  10337.00  14.206040
2018-01-17 17:21:00+00:00  10337.01  10337.01  10180.01  10192.15   7.906481
2018-01-17 17:22:00+00:00  10199.99  10200.00  10129.88  10160.08  28.119562
2018-01-17 17:23:00+00:00  10140.01  10161.00  10115.00  10115.01  11.283704
2018-01-17 17:24:00+00:00  10115.00  10194.99  10102.35  10194.99  10.617131
2018-01-17 17:25:00+00:00  10194.99  10240.00  10194.98  10220.00   8.586766
2018-01-17 17:26:00+00:00  10210.02  10210.02  10101.00  10138.00   6.616969
2018-01-17 17:27:00+00:00  10137.99  10138.00  10108.76  10124.94   9.962978
2018-01-17 17:28:00+00:00  10124.95  10142.39  10124.94  10142.39   2.262249

• write data

import numpy as np
import pandas as pd
data = np.array([(pd.Timestamp('2017-01-01 00:00').value / 10**9, 10.0)], dtype=[('Epoch', 'i8'), ('Ask', 'f4')])
cli.write(data, 'TEST/1Min/Tick')
# Out[10]: {'responses': None}

cli.query(pymkts.Params('TEST', '1Min', 'Tick')).first().df()

shows

                            Ask
Epoch
2017-01-01 00:00:00+00:00  10.0

• Variable length records

Marketstore is a database that achieves high performance by limiting the number of records in a timeframe to one. The supported timeframes are from 1D (1 day) to 1Sec (1 second), and basically, the longer the timeframe is, the faster you can read and write data.

However, it also supports data that does not arrive at a specific interval or more frequently than every second, such as board data and TICK data. Such kind of data is called variable-length records in marketstore.

You can use the variable-length records feature by specifying isvariablelength=True when you write with pymarketstore.

import numpy as np, pandas as pd, pymarketstore as pymkts

symbol, timeframe, attribute_group = "TEST", "1Sec", "Tick"
data_type = [('Epoch', 'i8'), ('Bid', 'f4'), ('Ask', 'f4'), ('Nanoseconds', 'i4')]
tbk = "{}/{}/{}".format(symbol, timeframe, attribute_group)
client = pymkts.Client()

# --- write variable-length records (=multiple records in a single timeframe (1Sec))
data = np.array([
    (pd.Timestamp('2021-01-01 00:00:00').value / 10 ** 9, 10.0, 20.0, 1000000),
    (pd.Timestamp('2021-01-01 00:00:00').value / 10 ** 9, 30.0, 40.0, 2000000),
    (pd.Timestamp('2021-01-01 00:00:00').value / 10 ** 9, 50.0, 60.0, 3000000),
], dtype=data_type)
client.write(data, tbk, isvariablelength=True)

# --- query variable-length records
params = pymkts.Params(symbol, timeframe, attribute_group)
print(client.query(params=params).first().df())

# --- tearDown
client.destroy(tbk)

shows

                            Bid Ask Nanoseconds
Epoch                                             
2021-01-01 00:00:00+00:00 10.0 20.0 1000000
2021-01-01 00:00:00+00:00 30.0 40.0 2000000
2021-01-01 00:00:00+00:00 50.0 60.0 3000000

Because the data type of Epoch column is always set as 'i8'(=int64) and it's not sufficient for describing a date with sub-second accuracy, the sub-second information is stored in another column (=Nanoseconds column, the data type is 'i4') in marketstore.

Command-line

Connect to a marketstore instance with

// For a local db-
marketstore connect --dir <path>
// For a server-
marketstore connect --url <address>

and run commands through the sql session.

Plugins

Go plugin architecture works best with Go1.10+ on linux. For more on plugins, see the plugins package Some featured plugins are covered here -

Streaming

You can receive realtime bars updates through the WebSocket streaming feature. The db server accepts a WebSocket connection on /ws, and we have built a plugin that pushes the data. Take a look at the package for more details.

GDAX Data Feeder

The batteries are included, so you can start pulling crypto price data from GDAX right after you install MarketStore. Then you can query DataFrame content over the network at as low latency as your local HDF5 files from disk, and appending new data to the end is two orders of magnitude faster than DataFrame would be. This is because the storage format is optimized for the type of data and use cases as well as for modern filesystem/hardware characteristics.

You can start pulling data from GDAX if you configure the data poller. For more information, see the package

On-Disk Aggregation

This plugin allows you to only worry about writing tick/minute level data. This plugin handles time-based aggregation on disk. For more, see the package

Replication

You can replicate data from a master marketstore instance to other marketstore instances. In mkts.yml config file, please set the config as the following:

• master instance

replication:
  # when enabled=true, this instance works as master instance and accept connections from replicas
  enabled: true
  # when tls_enabled=true, transport security between master and replica is enabled.
  # tls_enabled: true # both master and replica should have tls_enabled=true to enable TLS
  # public/private key pair from a pair of files. The files must contain PEM encoded data.
  # The cert file may contain intermediate certificates following the leaf certificate to form a certificate chain.
  # cert_file: "/Users/dakimura/projects/misks/tmpcert/server.crt" # both master and replica should have this config to enable TLS
  # key_file: "/Users/dakimura/projects/misks/tmpcert/server.key"
  # port to be used for the replication protocol
  listen_port: 5996

• replica instance(s)

replication:
  # when master_host is set, this instance works as a replica instance
  master_host: "127.0.0.1:5995"
  # when tls_enabled=true on master server, GRPC communication between master and replica is encrypted by SSL.
  # tls_enabled: true
  # cert_file: "/Users/dakimura/projects/misks/tmpcert/server.crt" # both master and replica should have this config to enable TLS

limitations

• Currently, the replication connection is initialized only at a startup of a marketstore replica instance. Please be sure to start the master instance first when you want to replicate data.

• Currently, only write API is supported. delete API result won't be reflected to replica instances.

• When replication is enabled on a replica instance, the instance is set to read-only mode and write API call(s) to the instance will fail.

Development

If you are interested in improving MarketStore, you are more than welcome! Just file issues or requests in GitHub or contact [email protected]. Before opening a PR please be sure tests pass-

make unittest

Plugins Development

We know the needs and requirements in this space are diverse. MarketStore provides strong core functionality with flexible plug-in architecture. If you want to build your own, look around plugins