Time series databases (TSDBs) are quite popular these days. To name a few, there are InfluxDB, Graphite, Druid, Kairos, and Prometheus. All aim to optimize data storage and querying for time-based data, which is highly relevant in a physics labs where there are multitude of "metrics" (to borrow a phrase used frequently in TSDB documentation) that naturally lend themselves to time series representation: lab (and individual device) temperatures, vacuum chamber pressures, and laser powers, just to name a few. Ideally, one could log various data to one of these databases and then use a tool like Grafana to visualize it. Sadly, more traditional relational databases like SQLite and PostgreSQL are not (currently) supported by Grafana (although this is now being addressed by a datasource plugin in development).

Nevertheless, there are quite a few reasons to favor a traditional RDBMS over a newfangled TSDB. To name a few:

  • Longevity: SQL has been around since the 1970s and became standardized in the 1980s.
  • Ubiquity: almost every server (web or otherwise) has an instance of SQL installed. If not, SQLite doesn't even require a server!
  • Community: not to suggest there aren't good communities with TSDBs, but the Postgres and SQLite communities in particular are generally quite helpful. Combined with the longevity aspect, any question one may have about how to accomplish a particular task with a SQL database is likely to be easily answerable with a simple web search.

In this post, I will outline a few things I have learned in using SQL for storing time series data. In particular, I will focus on Postgres, but the same general principles apply to other dialects. Sample code for some examples can be found on GitLab.

Schema definition

One "disadvantage" to SQL is it traditionally requires tightly defined schema. In practice when logging time series data, this is not usually a problem since each measurement device can neatly have its own column. Where this can become somewhat of nuiscance is when adding new devices. InfluxDB (for example) gets around this with its query language being quite flexible. In traditional SQL, the approach would require altering a table to add a new column. This is not too difficult in principle, but requires a (naive) program for logging data to frequently make ALTER TABLE calls and check if columns already exist. (Note that if using Python, this can be easily dealt with by using the dataset library.).

In real laboratories, though, we tend to know the kinds of things we are going to measure. So even if we add new devices that we want to log data from, we can still come up with a reasonable schema definition that fits well within the SQL paradigm. As an example, let's consider storing data from thermocouples in a table. We could get away with as few as three columns to describe the data: a timestamp (of course), a name or unique ID of the sensor, and a temperature measurement. For good measure, we should also add a primary key ID column to make a grand total of four columns. So far, our table looks like this:

 id |           timestamp           |  sensor   | temperature

For the timestamp column, I highly recommend using TIMESTAMP WITH TIME ZONE rather than TIMESTAMP WITHOUT TIME ZONE (more on why later).

For efficient querying, we'll want to index the timestamp and sensor columns. Depending on the number of sensors, it may also make sense to make a combined index on both, but we can defer this decision to later if it becomes necessary. Using SQLAlchemy, we define our table like this:

metadata = sa.MetaData()
table = sa.Table(
    'timeseries', metadata,
    sa.Column('id', sa.Integer, primary_key=True),
    sa.Column('timestamp', sa.DateTime(timezone=True),
              nullable=False, index=True),
    sa.Column('sensor', sa.String(length=128), nullable=False, index=True),
    sa.Column('temperature', sa.Float(precision=4), nullable=False))

which results in the following SQL:

CREATE TABLE timeseries (
    sensor VARCHAR(128) NOT NULL,
    temperature FLOAT(4) NOT NULL,
    PRIMARY KEY (id)
CREATE INDEX ix_timeseries_sensor ON timeseries (sensor);
CREATE INDEX ix_timeseries_timestamp ON timeseries (timestamp);

Basic querying

Simple queries are performed as normal:

SELECT * FROM timeseries WHERE sensor = 'sensor_01';

Postgres has quite a few date and time functions for building more complicated queries. It understands ISO 8601 out of the box:

test=> SELECT * FROM timeseries
test-> WHERE timestamp > '2016-06-13T22:00+02'
test-> AND sensor = 'sensor_01';
 id |           timestamp           |  sensor   | temperature
  8 | 2016-06-14 23:18:16.149606+02 | sensor_01 |     22.7061
  4 | 2016-06-14 23:18:11.985645+02 | sensor_01 |     25.4643
(2 rows)

Here we are explicit with the UTC offset of +2 hours (CEST). If omitted, the server locale is assumed. This brings us to why we should bother with time zones in the first place: internally, we want all timestamps stored in UTC to avoid ambiguity (Postgres already does this internally). Externally, (e.g., from Python scripts), we want to be able to use whatever time zone we're in to not have to think too hard.

SQLAlchemy treats naive datetime objects, uh, naively. This means that if a new datetime is created without explicitly specifying a time zone, that +2 hours above is lost and our time quries will start to get confusing. To avoid this problem, the best solution I have found is to always declare columns as TIMESTAMP WITH TIME ZONE (DateTime(timezone=True) in SQLAlchemy terms) and explicitly. So rather than inserting new timestamps with

from datetime import datetime
# ...
timestamp = datetime.now()

instead prefer

from datetime import datetime, timezone
# ...
timestamp = datetime.now(timezone.utc)

Aside: why oh why doesn't datetime.utcnow just do this?

Now we can build queries in Python like this using pandas and raw SQL queries:

today = datetime.now(timezone.utc)
today = today.replace(hour=0, minute=0, second=0, microsecond=0)
query = (
    "SELECT * FROM timeseries " +
    "WHERE timestamp >= '{}' ".format(today.isoformat()) +
    "AND sensor = 'sensor_01'"
df = pd.read_sql_query(query, engine, index_col="timestamp")

Data aggregation

Depending on data density, it may be useful to downsample data and look at aggregates such as the mean temperature in half-hour windows over the course of a day. We can easily accomplish this after the fact with pandas, but we can just as easily use Postgres aggregate functions to do this for us on the server. One advantage to this approach is a reduction in network overhead, which is especially relevant for very large datasets. Another is that these queries can be cached using materialized views. (This is a more advanced topic that I will not cover here. Instead, see the link in the references section below for a good treatment).

The key here is to use the date_trunc aggregate function and GROUP BY to only look at (for example) one hour at a time. An example of an aggregate query:

  date_trunc('hour', timestamp) AS timestamp,
  avg(temperature) AS temperature
FROM timeseries
WHERE timestamp >= '2016-06-25'
AND sensor = 'sensor_01'
GROUP BY date_trunc('hour', timestamp)
ORDER BY timestamp;

which results in something like:

timestamp        |   temperature    
2016-06-25 00:00:00+02 | 22.0828623312065
2016-06-25 01:00:00+02 | 22.0026334276975
2016-06-25 02:00:00+02 | 21.9871146672498
2016-06-25 03:00:00+02 | 22.0274553065207
2016-06-25 04:00:00+02 | 21.9357200048187
2016-06-25 05:00:00+02 | 21.9737668623899
2016-06-25 06:00:00+02 | 22.0098525849685
2016-06-25 07:00:00+02 | 22.0767008988982
2016-06-25 08:00:00+02 | 22.2146511332874
2016-06-25 09:00:00+02 | 21.9118559617263
2016-06-25 10:00:00+02 | 22.0417969508838
2016-06-25 11:00:00+02 | 22.0554379473676
2016-06-25 12:00:00+02 | 22.0193907419841
2016-06-25 13:00:00+02 | 22.0560295554413
2016-06-25 14:00:00+02 | 21.8087244594798
2016-06-25 15:00:00+02 | 22.0494429762518
2016-06-25 16:00:00+02 | 21.9082782661007
2016-06-25 17:00:00+02 | 21.4403478373652
(18 rows)

Other strategies

Another approach to avoid the time zone issue entirely is to simply store timestamps using something like UNIX time. Since pretty much every programming language imaginable has a built-in function to return time in seconds since the epoch, this is a reasonable approach (and is a bit more portable to other SQL dialects). The major downside to this is that compared to ISO 8601, UNIX time is not as readable by humans and therefore may require extra steps to convert to and from a human-readable format.

Depending on what you are doing with your data, it could also make sense to store, say, an hour's worth of data in a single row using the ARRAY data type. Combining arrays with array functions could then effectively do aggregation (somewhat) automatically rather than by query. This could also mean a bit of extra work when inserting new data or getting data stored in the database into a form friendly to your data analysis tools of choice.

References and further reading