tspDB enables predictive query functionality in PostgreSQL by building an additional “prediction index” for a collection of time-series of interest.

A prediction index will allow for fast data retrieval, but for entries that are:

1- At a future time step (i.e. forecasting);

2- Missing/corrupted by noise (i.e. imputation)

Our paper here provides more information about how tspDB works and its performance.  

This work has the following dependencies:

• PostgreSQL 12+
• Python 3.5+ with the libraries: (numpy, pandas, scipy, sklearn)

For insallation instruction, go to the installation page here

The main functionalities of tspDB is enabling predictive queries, which are enabled via creating a prediction index on your time series table. The index is created via the function create_pindex(). which you can use as follow:

select create_pindex('tablename','time_column_name','{"value_column_name"}','index_name')

To get you familiar with tspDB capabilities, we provided a testing function that will create a set of time series tables in your database. The test function will also create several prediction indices. Run the function from your postgres terminal

SELECT test_tspdb();

if you get at the last line

NOTICE:  Pindices successfully created

then the test has passed. Now we can check the predictive indices the test has created through

SELECT * FROM list_pindices();

You will see the three predictive indices created by the test. Now let's create our own predictive index, which we will call 'pindex1' on the time series table mixturets2. The prediction index is created on the time column time and the value column ts_7:

SELECT create_pindex('mixturets2','time','{"ts_7"}','pindex1');

we can see our newly created index by running list_pindices again:

SELECT * FROM list_pindices();

Let's now use that prediction index to produce some predictions! Let's for example predict at a time t that exists in the database. Effectively, we are denoising the existing observation or imputing a null observation. For example, at time 1, ts_7 has a null value as you can see by running:

SELECT ts_7 FROM mixturets2 WHERE time = 1;

Let's impute this point by running:

SELECT * FROM predict('mixturets2','ts_7',1,'pindex1');

Which will return predictions as well as upper an lower bound for a 95% confidence interval. We can get a tighter bound with lower confidence by changing the confidence interval to, say 80%:

SELECT * FROM predict('mixturets2','ts_7',1,'pindex1', c=> 80);

The prediction index also support forecasting queries using the same function. For example, you can forecast the value of column ts_7 at time 100010, ten points ahead of what exists in the database by running:

SELECT * FROM predict('mixturets2','ts_7',100010,'pindex1');

In a similar fashion, you can execute range predictive queries using predict(). For example, we can impute the first hundered points of ts_7 using:

SELECT * FROM predict('mixturets2','ts_7',0,100,'pindex1');

or forecast the next 10 points using:

SELECT * FROM predict('mixturets2','ts_7',100001,100010,'pindex1');

For further examples, check the python notebook examples here

Please visit our Github page for more information

This work is licensed under the Apache 2.0 License.