Machine learning driven data collection of high-frequency network telemetry for failure prediction

1. A method comprising:

receiving, at a supervisory service for one or more networks, telemetry data samples from a plurality of networking devices in the one or more networks;

training, by the service, a failure prediction model to predict failures in the one or more networks, using a training dataset comprising the received telemetry data samples;

assessing, by the service, performance of the failure prediction model;

training, by the service and based on the assessed performance of the failure prediction model, a machine learning-based classification model to determine whether a networking device should send a particular telemetry data sample to the service; and

sending, by the service, the machine learning-based classification model to one or more of the plurality of networking devices, to control which telemetry data samples the one or more networking devices send to the supervisory service.

2. The method as in claim 1 , wherein assessing performance of the failure prediction model comprises:

computing, for each feature of the training dataset, a measure of feature importance to the failure prediction model.

3. The method as in claim 2 , further comprising:

instructing, by the service, a particular one of the plurality of networking devices to stop reporting a particular feature to the supervisory service, based on its associated measure of feature importance to the failure prediction model.

4. The method as in claim 1 , further comprising:

determining, for each of the telemetry data samples of the training dataset, a measure of information gain that it provides to the failure prediction model.

5. The method as in claim 4 , wherein the machine learning-based classification model is trained to determine whether a networking device should send a particular telemetry data sample to the service, based on a predicted measure of information gain of the particular telemetry data sample.

6. The method as in claim 5 , wherein the machine learning-based classification model is trained to determine whether a networking device should send a particular telemetry data sample to the service, based further on a predicted resource cost of sending the particular telemetry data sample to the service.

7. The method as in claim 1 , further comprising:

training a plurality of failure prediction models to predict failures in the one or more networks, using telemetry data collected by the networking devices at different granularities;

comparing accuracies of the plurality of failure prediction models, to determine a granularity for telemetry data collection by the networking devices; and

instructing one or more of the networking devices to collect telemetry data at the determined granularity.

8. The method as in claim 1 , wherein assessing performance of the failure prediction model comprises:

deploying the failure prediction model to a particular one of the networking devices; and

receiving feedback regarding whether the deployed model correctly predicted failures.

9. The method as in claim 8 , further comprising:

instructing the particular networking device to increase its telemetry data sampling frequency, when the feedback indicates that the deployed model is inaccurate on the particular networking device.

10. The method as in claim 1 , wherein the plurality of networking devices comprise an edge router connected to a software-defined wide area network (SD-WAN).

11. An apparatus, comprising:

one or more network interfaces to communicate with one or more networks;

a processor coupled to the network interfaces and configured to execute one or more processes; and

a memory configured to store a process executable by the processor, the process when executed configured to:

receive telemetry data samples from a plurality of networking devices in the one or more networks;

train a failure prediction model to predict failures in the one or more networks, using a training dataset comprising the received telemetry data samples;

assess performance of the failure prediction model;

train, based on the assessed performance of the failure prediction model, a machine learning-based classification model to determine whether a networking device should send a particular telemetry data sample to the apparatus; and

send the machine learning-based classification model to one or more of the plurality of networking devices, to control which telemetry data samples the one or more networking devices send to the apparatus.

12. The apparatus as in claim 11 , wherein the apparatus assesses performance of the failure prediction model by:

computing, for each feature of the training dataset, a measure of feature importance to the failure prediction model.

13. The apparatus as in claim 12 , wherein the process when executed is further configured to:

instruct a particular one of the plurality of networking devices to stop reporting a particular feature to the apparatus, based on its associated measure of feature importance to the failure prediction model.

14. The apparatus as in claim 11 , wherein the process when executed is further configured to:

determine, for each of the telemetry data samples of the training dataset, a measure of information gain that it provides to the failure prediction model.

15. The apparatus as in claim 14 , wherein the machine learning-based classification model is trained to determine whether a networking device should send a particular telemetry data sample to the apparatus, based on a predicted measure of information gain of the particular telemetry data sample.

16. The apparatus as in claim 15 , wherein the machine learning-based classification model is trained to determine whether a networking device should send a particular telemetry data sample to the apparatus, based further on a predicted resource cost of sending the particular telemetry data sample to the apparatus.

17. The apparatus as in claim 11 , wherein the process when executed is further configured to:

train a plurality of failure prediction models to predict failures in the one or more networks, using telemetry data collected by the networking devices at different granularities;

compare accuracies of the plurality of failure prediction models, to determine a granularity for telemetry data collection by the networking devices; and

instruct one or more of the networking devices to collect telemetry data at the determined granularity.

18. The apparatus as in claim 11 , wherein the apparatus assesses performance of the failure prediction model by:

deploying the failure prediction model to a particular one of the networking devices; and

receiving feedback regarding whether the deployed model correctly predicted failures.

19. The apparatus as in claim 18 , wherein the process when executed is further configured to:

instruct the particular networking device to increase its telemetry data sampling frequency, when the feedback indicates that the deployed model is inaccurate on the particular networking device.

20. A tangible, non-transitory, computer-readable medium storing program instructions that cause a supervisory service to execute a process comprising:

receiving, at a supervisory service for one or more networks, telemetry data samples from a plurality of networking devices in the one or more networks;

training, by the service, a failure prediction model to predict failures in the one or more networks, using a training dataset comprising the received telemetry data samples;

assessing, by the service, performance of the failure prediction model;

training, by the service and based on the assessed performance of the failure prediction model, a machine learning-based classification model to determine whether a networking device should send a particular telemetry data sample to the service; and

sending, by the service, the machine learning-based classification model to one or more of the plurality of networking devices, to control which telemetry data samples the one or more networking devices send to the supervisory service.