RedisTimeseriesManager is a redis timeseries management system that enhance redis timeseries with features including multi-line data, built-in timeframes, data classifiers and convenient data accessors. This is achieved by maintaing a set of timeseries that are tied together(called lines) and interact with them as a whole.
As a result, multiple timeseries values can be refered with a timestamp as if they are stored in a table with a timestamp and multiple columns. RedisTimeseriesManager
uses RedisTimeSeries to store timeseries data.
This library supports classification of data at multiple levels, plus support for timeframes and data compression(downsampling).
To install RedisTimeseriesManager, run the following command:
pip install --upgrade redis_timeseries_manager
If you are upgrading from an older version, see the change log
To get started, simply create a class that inherits from RedisTimeseriesManager
. Then set the properties _name
, _lines
, and _timeframes
.
from redis_timeseries_manager import RedisTimeseriesManager
class Test(RedisTimeseriesManager):
_name = 'test'
_lines = ['l1', 'l2']
_timeframes = {
'raw': {'retention_secs': 60*60*24}, # retention 1 day
}
You can think of _lines
as the columns in a relational database. You can have as many lines as your requirements. They can be added or removed at any time using add_line()
or delete_line()
methods.
At least one timeframe must be provided. Even if you simply want to store timeseries data without separate timeframes, add a default timeframe and it will be used seamlessly. Also a unique _name
must be provided for each class.
t = Test(host, port, db, password)
The method insert()
is used to add data to the series. The syntax is:
t.insert(data, c1, c2, create_inplace=False)
The format of data is as follows:
[timestamp, l1, l2, ...]
c1
and c2
are the classifiers. You can classify data in two levels using these classifiers. In our sensors example we have used c1
to identify the location of sensor and c2
for the sensor itself.
Before adding data, we have to prepare the timeseries with the classifiers. To achieve this, we use the create()
method. For the sensors example:
t.create(
c1='building1',
c2='sensor1',
)
Now we can add data for sensor 1:
t.insert(
data=[
[123456, 1, 2],
[123457, 3, 4],
[123458, 5, 6],
],
c1='building1',
c2='sensor1',
)
For the sensor2, we don't want to prepare series explicitly, instead we set create_inplace
to True and the series will be prepared with the new classifiers implicitely while inserting data:
t.insert(
data=[
[123456, 7, 8],
[123457, 9, 10],
[123458, 11, 12],
],
c1='building1',
c2='sensor2',
create_inplace=True,
)
The method read()
is used to read data from the series. In our example we can read the data for sensor1 as follows:
t.read(
c1='building1',
c2='sensor1',
)
# For convenience, an alternate way to provide classifiers while reading the data is the filters paramter:
t.read(
filters={
'c1': 'building1',
'c2': 'sensor1',
}
)
[[123456, 1.0, 2.0], [123457, 3.0, 4.0], [123458, 5.0, 6.0]]
There are also some other parameters in the read()
method that can help further filter out data stored in the timeseries or to investigate about them. They include timeframe
, from_timestamp
, to_timestamp
, limit
, read_from_last
and latest
. You can also modify the output format using line_order
and return_as
parameters. Refer to the corresponding documentation in the read()
method for details.
Since v2.0, pandas dataframes are supported. You can choose the format of output data when calling read methods. Supported formats are 'list' (python list
, default format), 'df'(pandas dataframe), 'dict' (dictioanry of lines), 'sets-list' (sets of data including lables and lists), 'sets-df' (sets of data including labels and dataframes), 'sets-dict' (sets of data including labels and dictionary of lines) and 'raw' which is the raw data read from the timeseries.
Lines data at an existing timestamp can be updated individually by taking advantage of update(values:dict, c1:str, c2:str, timestamp:int)
method.
values
are key-value pairs of data to be updated at the time timestamp
and all provided keys
must correspond to an existing line
in timeseries. If you don't include a line, the value at that line will be untouched.
Also the update will be applied on the first timeframe and the other timeframes will be updated by compaction rules(if any).
Usage example for update:
t.update(
values={
'l2': 100,
},
c1='building1',
c2='sensor2',
timestamp=123457
)
If you do not need data to be compressed across timeframes, you can set only a single timeframe in _timeframes
class property. This will fully disable compaction functionality; but note that at least one timeframe must be set always.
WARNING: Due to an unfixable bug in redis timeseries module only use
db
with index0
while data compaction is required; otherwise compaction rules won't work.
To have a separate timeframe without data compaction, set ignore_rules
to True
in the timeframe definition:
_timeframes = {
'1m': {'retention_secs': 60*60*24*10},
'1h': {'retention_secs': 60*60*24*90, 'bucket_size_secs': 3600},
'1d': {'retention_secs': 60*60*24*365, 'ignore_rules': True},
}
In the above example, the 1d
timeframe is isolated and no compaction rule will have interaction with that. Data can be inserted into this timeframe using insert(c1=..., timeframe='1d')
One usage may be in the case that you want to keep track and maintain the minute data but have a separate data source for daily data. Keep in mind that you should never write data directly into the timeframes that the result of compaction rules are written. In the above example, the 1m
(default) and 1d
timeframes are safe to write directly.
While in most use cases, two classifiers for the data must be enough; there might be scenarios where more than two classifiers for the data is required. In such cases, you can extend the classifiers in c1
or c2
classifier.
As of version 2.1, redis_timeseries_manager supports extra_labels
that gives the ability to set custom labels for the data. The main advantage of labels in redis timeseries emerges when you utilize them with redis multi-timeseries commands like TS.MRANGE
The extending process consist of two parts: First we have to provide a unique identifier as the classifier and secondary provide the corresponding labels that identify the data as extra_labels
. (Do not include timeframe in this process, timeframes are fully handled internally)
For better clarification, suppose a scenaro where we are required to store performance of several users who are optimizing strategies on given sample data. In this case we need 4 different classifiers and we have to extend additional ones in a classifier like c2
.
Here is the full example:
from redis_timeseries_manager import RedisTimeseriesManager
class Measurements(RedisTimeseriesManager):
_name = 'feature_tests'
_lines = ['l1', 'l2']
_timeframes = {
'raw': {'retention_secs': 100000}
}
settings = {
'host': 'localhost',
'port': 6379,
'db': 0,
'password': None,
}
tl = Measurements(**settings)
tl.insert(
data=[
[123456, 7, 8],
[123457, 9, 10],
[123458, 11, 12],
[123459, 13, 14],
],
c1='performance',
c2='u_1_22_46', # generally you have to generate this string programmatically
extra_labels={
'user_id': 1,
'strategy_id': 22,
'sample_id': 46
},
create_inplace=True,
)
tl.insert(
data=[
[123456, 17, 18],
[123457, 19, 110],
[123458, 111, 112],
[123460, 113, 114],
],
c1='performance',
c2='u_2_22_46',
extra_labels={
'user_id': 2,
'strategy_id': 22,
'sample_id': 46
},
create_inplace=True,
)
Later, to read data, we have to provide the full labels we have decided to define(and differentiate) the data with, in place of using a plain c2
classifier:
tl.read(
c1='performance',
c2={
'user_id': 2,
'strategy_id': 22,
'sample_id': 46
},
return_as='df'
)
# the same functionality can be achieved using the filters parameter:
tl.read(
filters={
'c1': 'performance',
'user_id': 2,
'strategy_id': 22,
'sample_id': 46
}
return_as='df'
)
time | l1 | l2 |
---|---|---|
123456 | 17.0 | 18.0 |
123457 | 19.0 | 10.0 |
123458 | 111.0 | 112.0 |
123460 | 113.0 | 114.0 |
If we don't provide the full labels, multiple data-points with the same time
might return and this is usually not we expect from a timeseries data and that's why RedisTimeseriesManager
by default prevents this to happen. However if you persist, you can turn the allow_multiple
option on to let multiple data sets to be combined together.
In our exampe, if you wanted all the entries for the strategy_id
of 22, you can do as following:
tl.read(
c1='performance',
c2={
'strategy_id': 22,
},
allow_multiple=True,
return_as='df'
)
time | l1 | l2 |
---|---|---|
123456 | 7.0 | 8.0 |
123456 | 17.0 | 18.0 |
123457 | 9.0 | 10.0 |
123457 | 19.0 | 110.0 |
123458 | 11.0 | 12.0 |
123458 | 111.0 | 112.0 |
123459 | 13.0 | 14.0 |
123460 | 113.0 | 114.0 |
WARNING: Consider only concepts that distinguish/identify the class of data as the extended classifiers. This means that the number of these concepts will be very few while many other concepts can be considered in the form of data.
In this example, we are going to maintain data of two imaginary sensors. Each sensor provides two measurements: temperature
and humidity
.
The data is collected with the resolution of one minute. Then we compress(downsample) the data to hourly and daily resolutions. To compress the data, we consider the average value of temperature and the maximum value of humidity in each time frame.
We also want to keep 1-minute sensor data for just one week, 1-hour data for one month and respectively 1-day data for a year. In this Example, we use the classifier 1(c1) to identify the building where the sensor is located and the classifier 2(c2) for the sensor.
import time, datetime, random
from pytz import timezone
from redis_timeseries_manager import RedisTimeseriesManager
settings = {
'host': 'localhost',
'port': 6379,
'db': 13,
'password': None,
}
class SensorData(RedisTimeseriesManager):
_name = 'sensors'
_lines = ['temp', 'hum']
_timeframes = {
'raw': {'retention_secs': 60*60*24*7}, # retention 7 day
'1h': {'retention_secs': 60*60*24*30, 'bucket_size_secs': 60*60}, # retention 1 month; timeframe 3600 secs
'1d': {'retention_secs': 60*60*24*365, 'bucket_size_secs': 60*60*24}, # retention 1 year; timeframe 86400 secs
}
#compaction rules
def _create_rule(self, c1:str, c2:str, line:str, timeframe_name:str, timeframe_specs:str, source_key:str, dest_key:str):
if line == 'temp':
aggregation_type = 'avg'
elif line == 'hum':
aggregation_type = 'max'
bucket_size_secs = timeframe_specs['bucket_size_secs']
self._set_rule(source_key, dest_key, aggregation_type, bucket_size_secs)
@staticmethod
def print_data(data):
for ts, temp, hum in data:
print(f"{datetime.datetime.fromtimestamp(ts, tz=timezone('UTC')):%Y-%m-%d %H:%M:%S}, temp: {round(temp, 2)}, hum(max): {round(hum, 2)}")
Demonstrating a high-performance market price data downsampling mechanism using redis backend and RedisTimeseriesManager
In this example, we are going to maintain the data of some financial markets. We have chosen the cryptocurrency
and irx
for our example. Each market contain several instruments that we refer to them as symbols and we collect OHLCV(open
, high
, low
, close
, volume
) data for each symbol.
The raw data is directly collected from the market with the resolution of seconds and we insert them in raw
timeframe. Then we compress(downsample) the data to timeframes of 1m
, 1h
and 1d
. As the names open
, high
, low
, close
, volume
implies, we use the FIRST
aggregator for open
, MAX
for high
, MIN
for low
, LAST
for close
and the SUM
aggregator for volume
to compress the data and build the appropriate timeframes of data.
We also want to keep 1m
data for just one week, 1h
for one month and respectively 1d
data for a year.
In this Example, we use the classifier 1(c1) to identify the market(here cryptocurrency
or irx
) and the classifier 2(c2) for the symbols.
import time, datetime, random
from pytz import timezone
from redis_timeseries_manager import RedisTimeseriesManager
settings = {
'host': 'localhost',
'port': 6379,
'db': 13,
'password': None,
}
class MarketData(RedisTimeseriesManager):
_name = 'markets'
_lines = ['open', 'high', 'low', 'close', 'volume']
_timeframes = {
'raw': {'retention_secs': 60*60*24*4}, # retention 4 days
'1m': {'retention_secs': 60*60*24*7, 'bucket_size_secs': 60}, # retention 7 day; timeframe 60 secs
'1h': {'retention_secs': 60*60*24*30, 'bucket_size_secs': 60*60}, # retention 1 month; timeframe 3600 secs
'1d': {'retention_secs': 60*60*24*365, 'bucket_size_secs': 60*60*24}, # retention 1 year; timeframe 86400 secs
}
#compaction rules
def _create_rule(self, c1:str, c2:str, line:str, timeframe_name:str, timeframe_specs:str, source_key:str, dest_key:str):
if line == 'open':
aggregation_type = 'first'
elif line == 'close':
aggregation_type = 'last'
elif line == 'high':
aggregation_type = 'max'
elif line == 'low':
aggregation_type = 'min'
elif line == 'volume':
aggregation_type = 'sum'
else:
return
bucket_size_secs = timeframe_specs['bucket_size_secs']
self._set_rule(source_key, dest_key, aggregation_type, bucket_size_secs)
@staticmethod
def print_data(data):
for ts, open, high, low, close, volume in data:
print(f"{datetime.datetime.fromtimestamp(ts, tz=timezone('UTC')):%Y-%m-%d %H:%M:%S}, open: {open}, high: {high}, low: {low}, close: {close}, volume: {volume}")
(Since version 2.3)
- In favor of improvements in read() method, some of the old methods associated with reading information have become obsolete. They include read_last_n_records(), read_last_nth_record() and read_last(). Similar functionality is available in read() method.
- The read() method does not rely on classifier(c1 and c2) parameters anymore. You can now fully omit these parameters while reading data and instead use filters parameter. The filters parameter is all inclusive, meaning it can take c1, c2 (filters={'c1': 'something', 'c2': 'another'}) or in case of an extended classifier it takes label-value pairs of classifiers.