RedisTimeSeries 1.6 Is Out!

Today, we’re happy to announce the General Availability of RedisTimeSeries 1.6. This blog post details the major new features that are now available.

About RedisTimeSeries

RedisTimeSeries is a high-performance, memory-first time-series data structure for Redis. RedisTimeSeries supports time series multi-tenancy (it can hold many time series simultaneously) and can serve multiple clients accessing these time series simultaneously. It is now available also as part of Redis Stack.

Major new features in RedisTimeSeries 1.6

Enhanced querying capabilities
- Cross-time series aggregations
- Filtering results by timestamp or sample value
- Selecting which labels to retrieve
- Aligning aggregation buckets
  Deleting samples
Improved performance
Keyspace notifications
Redis Enterprise support for Redis on Flash

Enhanced Querying Capabilities

Cross-time series aggregations

Before version 1.6, only one type of aggregation was possible:

For a single time series –
aggregating the samples over equidurational timeframes.

Since version 1.6, two new aggregation types are possible:

For multiple time series –
aggregating samples with equal timestamps over the set of time series.
For multiple time series –
First, aggregate each time series over equidurational timeframes, then, for each timeframe, aggregate the results over the set of time series.

Let’s demonstrate the first new aggregation type. First, let’s create two stocks, and add their prices at 3 different timestamps:

Now, we can retrieve the maximal stock price per timestamp:
redis:6379> TS.MRANGE - + WITHLABELS FILTER type=stock GROUPBY type REDUCE max

1) 1) "type=stock" 2) 1) 1) "type" 2) "stock" 2) 1) "__reducer__" 2) "max" 3) 1) "__source__" 2) "stock:A,stock:B" 3) 1) 1) (integer) 1000 2) 120 2) 1) (integer) 1010 2) 110 3) 1) (integer) 1020 2) 120

The FILTER type=stock clause leaves us only with a single time series representing stock prices. The GROUPBY type REDUCE max clause splits the time series into groups with identical type values, and then, for each timestamp, aggregates all series that share the same type value using the max aggregator.

FILTER label=value is supported by TS.MRANGE and TS.MREVRANGE. Additional filtering clauses are supported as well (see documentation).

Next, we will demonstrate the second new aggregation type:

First, let’s create two stocks, and add their prices at 9 different timestamps.

TS.CREATE stock:A LABELS type stock name A TS.CREATE stock:B LABELS type stock name B TS.MADD stock:A 1000 100 stock:A 1010 110 stock:A 1020 120 TS.MADD stock:B 1000 120 stock:B 1010 110 stock:B 1020 100 TS.MADD stock:A 2000 200 stock:A 2010 210 stock:A 2020 220 TS.MADD stock:B 2000 220 stock:B 2010 210 stock:B 2020 200 TS.MADD stock:A 3000 300 stock:A 3010 310 stock:A 3020 320 TS.MADD stock:B 3000 320 stock:B 3010 310 stock:B 3020 300

Now, for each stock, we will calculate the average stock price per 1000 millisecond time frame, and then retrieve the stock with the

maximal average for that timeframe:

redis:6379> TS.MRANGE - + WITHLABELS AGGREGATION avg 1000 FILTER type=stock GROUPBY type REDUCE max

1) 1) "type=stock" 2) 1) 1) "type" 2) "stock" 2) 1) "__reducer__" 2) "max" 3) 1) "__source__" 2) "stock:A,stock:B" 3) 1) 1) (integer) 1000 2) 110 2) 1) (integer) 2000 2) 210 3) 1) (integer) 3000 2) 310

GROUPBY label REDUCE reducer is supported by TS.MRANGE and TS.MREVRANGE.

Reducer can be sum, min, or max.

Filtering results by timestamp or sample value

When using TS.RANGE, TS.REVRANGE, TS.MRANGE, and TS.MREVRANGE, you don’t always want to retrieve or aggregate all the samples.

Using [FILTER_BY_TS ts...] you can filter the sample by a list of specific timestamps.

Using [FILTER_BY_VALUE min max] you can filter the sample by minimal and maximal values.

Consider, for example, a sampled metric where the normal values are between -100 and 100, but the value 9999 is used as an indication of bad measurement.

TS.CREATE temp:TLV LABELS type temp location TLV TS.MADD temp:TLV 1000 30 temp:TLV 1010 35 temp:TLV 1020 9999 temp:TLV 1030 40

Now, let’s retrieve all values, ignore out-of-range values:

TS.RANGE temp:TLV - + FILTER_BY_VALUE -100 100

Let’s also retrieve the average value, ignore out-of-range values:

TS.RANGE temp:TLV - + FILTER_BY_VALUE -100 100 AGGREGATION avg 1000

Selecting which labels to retrieve

When using TS.MRANGE, TS.MREVRANGE, and TS.MGET, we don’t always want the values of all the labels associated with the matching time series, but only the values of selected labels.

SELECTED_LABELS allows selecting which labels to retrieve. Given the following time series and data:

TS.CREATE temp:TLV LABELS type temp location TLV TS.MADD temp:TLV 1000 30 temp:TLV 1010 35 temp:TLV 1020 9999 temp:TLV 1030 40

To get all the labels associated with matched time series we’ll use WITHLABELS:

redis:6379> TS.MGET WITHLABELS FILTER type=temp

1) 1) "temp:TLV" 2) 1) 1) "type" 2) "temp" 2) 1) "location" 2) "TLV" 3) 1) (integer) 1030 2) 40

But suppose we want only the location, we can use SELECTED_LABELS:

redis:6379> TS.MGET SELECTED_LABELS location FILTER type=temp

1) 1) "temp:TLV" 2) 1) 1) "location" 2) "TLV" 3) 1) (integer) 1030 2) 40

Aligning aggregation buckets

Suppose we want to receive the average daily temperature, but our ‘day’ starts at 06:00. In this case, we would like to ALIGN the intervals to 06:00 to 05:59, instead of to 00:00 to 23:59.

When using TS.RANGE, TS.REVRANGE, TS.MRANGE, and TS.MREVRANGE, it is now possible to align the aggregation buckets with the requested start, end, or specific timestamp – using ALIGN.

To demonstrate alignment, let’s add the following data:

TS.CREATE stock:A LABELS type stock name A TS.MADD stock:A 1000 100 stock:A 1010 110 stock:A 1020 120 TS.MADD stock:A 1030 200 stock:A 1040 210 stock:A 1050 220 TS.MADD stock:A 1060 300 stock:A 1070 310 stock:A 1080 320

Next, we’ll aggregate without using ALIGN (which means default alignment: 0)

redis:6379> TS.RANGE stock:A - + AGGREGATION min 20

1) 1) (integer) 1000 2) 100 2) 1) (integer) 1020 2) 120 3) 1) (integer) 1040 2) 210 4) 1) (integer) 1060 2) 300 5) 1) (integer) 1080 2) 320

And now with ALIGN:

redis:6379> TS.RANGE stock:A – + ALIGN 10 AGGREGATION min 20

1) 1) (integer) 990 2) 100 2) 1) (integer) 1010 2) 110 3) 1) (integer) 1030 2) 200 4) 1) (integer) 1050 2) 220 5) 1) (integer) 1070 2) 310

Setting ALIGN to 10 means that a bucket should start at time 10, and all the buckets (each with a 20 milliseconds duration) are aligned accordingly.

When the start timestamp for the range query is explicitly stated (not ‘-‘), it is also possible to set ALIGN to that time by setting align to ‘-‘ or to ‘start‘.

redis:6379> TS.RANGE stock:A 5 + ALIGN – AGGREGATION min 20

1) 1) (integer) 985 2) 100 2) 1) (integer) 1005 2) 110 3) 1) (integer) 1025 2) 200 4) 1) (integer) 1045 2) 220 5) 1) (integer) 1065 2) 310

Similarly, when the end timestamp for the range query is explicitly stated (not ‘+’), it is also possible to set ALIGN to that time by setting align to ‘+’ or to ‘end’.

Deleting samples

TS.DEL allows deleting samples in a given time series within two timestamps.

For example, TS.DEL stock:A 1020 1050 will delete all samples with timestamps between 1020 and 1050 (inclusive). The returned value is the number of samples deleted.

Improved performance

Many optimizations were implemented, and most queries would now execute much faster compared to RedisTimeSeries 1.4.

The following table details the number of queries per second achievable on a single node, for TSBS queries (which we described here). The table lists only the subset of the TSBS queries that were supported in version 1.4.

Query type	1.4 queries/sec	v1.6 queries/sec	% change (higher-better)
tsbs-scale100_cpu-max-all-1	1388	1500	8.07%
tsbs-scale100_double-groupby-1	100	108	8.00%
tsbs-scale100_groupby-orderby-limit	793	1282	61.66%
tsbs-scale100_single-groupby-1-1-1	13448	21347	58.74%
tsbs-scale100_single-groupby-1-1-12	2383	3921	64.54%

We can observe an improvement of 8% to 65% in the number of queries per second, compared to RedisTimeSeries 1.4.

Keyspace notifications

Redis Keyspace notifications allow Redis clients to subscribe to Pub/Sub channels in order to receive events affecting the Redis data set in some way. You can, for example, use RedisGears to trigger a function with these notifications.

As an example, it is possible to implement a time series predictor or anomaly detector that listens to the stream of samples and generates real-time predictions and warnings.
Check, for example, this test which subscribes to various RedisTimeSeries commands and generates events.

Redis Enterprise support for Redis on Flash

Since version 1.6, RedisTimeSeries can run in a Redis on Flash, but it’s important to note that RoF is implemented on the key level. Namely, the value of the whole time series lives either on FLASH or RAM.

RedisTimeSeries is part of Redis Stack

RedisTimeSeries is now part of Redis Stack. You can download the latest Redis Stack Server binaries for macOS, Ubuntu, or Redhat, or install them with Docker, Homebrew, or Linux.

Visualize time series data with RedisInsight

RedisInsight is a visual tool for developers that provides an excellent way to explore the data from RedisTimeSeries during development using Redis or Redis Stack.

You can execute time series queries and observe the results directly from the graphical user interface. RedisInsight can now visualize RedisTimeSeries query results.