Skip to content

Time and continuous aggregates

Learn to work with timezones and continuous aggregates

In TimescaleDB v2.20.0 and higher, a continuous aggregate that uses a function depending on the session timezone is created with a warning. Its results can change on rematerialization, because the setting changes from user to user. Inside time_bucket, and in earlier versions, such a function is an error.

To manage this, you can use explicit timezones in the view definition. Alternatively, you can create your own custom aggregation scheme for tables that use an integer time column.

The most common method of working with timezones is to declare an explicit timezone in the view query:

CREATE MATERIALIZED VIEW device_summary
WITH (timescaledb.continuous)
AS
SELECT
time_bucket('1 hour', observation_time) AS bucket,
min(observation_time AT TIME ZONE 'EST') AS min_time,
device_id,
avg(metric) AS metric_avg,
max(metric) - min(metric) AS metric_spread
FROM
device_readings
GROUP BY bucket, device_id;

Alternatively, store the timestamptz value and convert it when you query the view:

CREATE MATERIALIZED VIEW device_summary_tz
WITH (timescaledb.continuous)
AS
SELECT
time_bucket('1 hour', observation_time) AS bucket,
min(observation_time) AS min_time,
device_id
FROM
device_readings
GROUP BY bucket, device_id;
SELECT min_time::timestamp FROM device_summary_tz;

The cast uses the timezone of the session that runs the query.

Date and time is usually expressed as year-month-day and hours:minutes:seconds. Most TimescaleDB databases use a PostgreSQL date/time-type column to express the date and time. However, in some cases, you might need to convert these common time and date formats to a format that uses an integer. The most common integer time is Unix epoch time, which is the number of seconds since the Unix epoch of 1970-01-01, but other types of integer-based time formats are possible.

These examples use a hypertable called devices that contains CPU and disk usage information. The devices measure time using the Unix epoch.

For an integer-based time column, set tsdb.chunk_interval in the units of the column. In this case, each chunk covers one day.

Define the integer-based time column and chunk time interval:

CREATE TABLE devices(
time BIGINT, -- Time in seconds since epoch
cpu_usage INTEGER, -- Total CPU usage
disk_usage INTEGER, -- Total disk usage
PRIMARY KEY (time)
) WITH (
tsdb.hypertable,
tsdb.partition_column='time',
tsdb.chunk_interval='86400'
);

For TimescaleDB v2.23.0 and higher, the table is automatically partitioned on the first column in the table with a timestamp data type. To partition on another column, set tsdb.partition_column. If the table has no timestamp column, TimescaleDB throws an error and asks for an explicit definition. For earlier versions, set tsdb.partition_column to a time column.

If you are self-hosting TimescaleDB v2.20.0 to v2.22.1, to convert your data to the columnstore after a specific time interval, you have to call add_columnstore_policy after you call CREATE TABLE

If you are self-hosting TimescaleDB v2.19.3 and below, create a PostgreSQL relational table, then convert it using create_hypertable. You then enable hypercore with a call to ALTER TABLE.

To define a continuous aggregate on a hypertable that uses integer-based time, you need to have a function to get the current time in the correct format, and set it for the hypertable. You can do this with the set_integer_now_func function. It can be defined as a regular PostgreSQL function, but needs to be STABLE, take no arguments, and return an integer value of the same type as the time column in the table. When you have set up the time-handling, you can create the continuous aggregate.

  1. Set up the time conversion function
    CREATE FUNCTION current_epoch() RETURNS BIGINT
    LANGUAGE SQL STABLE AS $$
    SELECT EXTRACT(EPOCH FROM CURRENT_TIMESTAMP)::bigint;$$;
    SELECT set_integer_now_func('devices', 'current_epoch');
  2. Insert some rows
    CREATE EXTENSION tablefunc;
    INSERT INTO devices(time, cpu_usage, disk_usage)
    SELECT time,
    normal_rand(1,70,10) AS cpu_usage,
    normal_rand(1,2,1) * (row_number() over()) AS disk_usage
    FROM generate_series(1,10000) AS time;

    This command uses the tablefunc extension to generate a normal distribution, and uses the row_number function to turn it into a cumulative sequence.

  3. Create the continuous aggregate for the devices table
    CREATE MATERIALIZED VIEW devices_summary
    WITH (timescaledb.continuous) AS
    SELECT time_bucket('500', time) AS bucket,
    avg(cpu_usage) AS avg_cpu,
    avg(disk_usage) AS avg_disk
    FROM devices
    GROUP BY bucket;
  4. Check the view data
    SELECT * FROM devices_summary ORDER BY bucket LIMIT 10;

    You see something like:

    bucket | avg_cpu | avg_disk
    --------+---------------------+-----------------------
    0 | 69.8977955911823647 | 510.9739478957915832
    500 | 69.6340000000000000 | 1586.3920000000000000
    1000 | 70.4360000000000000 | 2489.7040000000000000
    1500 | 69.5820000000000000 | 3424.7480000000000000
    2000 | 69.7120000000000000 | 4382.1320000000000000
    2500 | 69.7220000000000000 | 5596.9300000000000000
    3000 | 69.6380000000000000 | 6453.6740000000000000
    3500 | 70.6200000000000000 | 7260.8820000000000000
    4000 | 70.1320000000000000 | 8611.2260000000000000
    4500 | 69.7580000000000000 | 9362.2100000000000000
    (10 rows)