Anomaly detection and filtering based on system logs

1. A method, comprising:

receiving, at a first time, a first sequence of log entries from a networked system;

generating, using a machine learning model and based on previously observed log entries including the first sequence of log entries, predicted log entries for a second time;

receiving, at the second time, a second sequence of log entries from the networked system;

determining whether the second sequence is anomalous based on comparing the second sequence with the predicted log entries;

in response to determining that the second sequence is anomalous, determining whether the second sequence is noteworthy based on a function entropy of the first sequence and a sentiment polarity of the second sequence, wherein the function entropy is a measure of randomness in the log entries, and wherein the sentiment polarity is a measure of sentiment of words found in the log entries; and

in response to determining that the second sequence is noteworthy, generating an anomaly report that includes the second sequence and a root cause.

2. The method of claim 1 , further comprising:

training the machine learning model based on non-anomalous log entries for the networked system;

maintaining a plurality of log templates for the log entries, wherein the plurality of log templates is updated during operation of the machine learning model to include new log templates observed in actual log templates from the networked system; and

adding a given log template to the plurality of templates after the given log template is observed in the log entries a threshold number of times.

3. The method of claim 2 , wherein the machine learning model is an LSTM auto-encoder operating with an attention layer that is fitted with unlabeled operational logs from the networked system, wherein the attention layer provides a vector representation based on outputs of the LSTM auto-encoder.

4. The method of claim 2 , wherein the predicted log entries are a set of log entries from the plurality of log templates with a highest probability based on the first sequence.

5. The method of claim 1 , wherein the second sequence is marked as not noteworthy when the function entropy is above an entropy threshold.

6. The method of claim 1 , wherein the second sequence is marked as not noteworthy when the sentiment polarity is above a positive sentiment threshold.

7. The method of claim 1 , wherein the root cause identifies log entries from the first sequence related to the second sequence and the anomaly report ranks identified anomalies based on the function entropy of the first sequence and the sentiment polarity of the second sequence.

8. A system, comprising:

a processor; and

a memory including instructions that when executed by the processor perform an operation comprising:

receiving, at a first time, a first sequence of log entries from a networked system;

generating, using a machine learning model and based on previously observed log entries including the first sequence of log entries, predicted log entries for a second time;

receiving, at the second time, a second sequence of log entries from the networked system;

determining whether the second sequence is anomalous based on comparing the second sequence with the predicted log entries;

in response to determining that the second sequence is anomalous, determining whether the second sequence is noteworthy based on a function entropy of the first sequence and a sentiment polarity of the second sequence, wherein the function entropy is a measure of randomness in the log entries, and wherein the sentiment polarity is a measure of sentiment of words found in the log entries; and

in response to determining that the second sequence is noteworthy, generating an anomaly report that includes the second sequence and a root cause.

9. The system of claim 8 , wherein a machine learning model determines that the second sequence is anomalous, wherein the machine learning model is an LSTM auto-encoder operating with an attention layer that is fitted with unlabeled previously observed log entries, wherein the attention layer provides a vector representation based on outputs of the LSTM auto-encoder.

10. The system of claim 9 , wherein a plurality of log templates is updated during operation of the machine learning model to include new log templates observed in actual log templates, the operation further comprising:

adding a given log template to the plurality of log templates after the given log template is observed in the log entries a threshold number of times.

11. The system of claim 8 , wherein the predicted log entries are a set of log entries from a plurality of log templates with a highest probability based on the first sequence.

12. The system of claim 8 , wherein the second sequence is marked as not noteworthy when the function entropy is above an entropy threshold.

13. The system of claim 8 , wherein the second sequence is marked as not noteworthy when the sentiment polarity is above a positive sentiment threshold.

14. The system of claim 8 , wherein the root cause identifies log entries from the first sequence related to the second sequence and the anomaly report ranks identified anomalies based on the function entropy of the first sequence and the sentiment polarity of the second sequence.

15. A device including computer-readable instructions that when executed by a processor in the device enable performance of an operation comprising:

receiving, at a first time, a first sequence of log entries from a networked system;

generating, using a machine learning model and based on previously observed log entries including the first sequence of log entries, predicted log entries for a second time;

receiving, at the second time, a second sequence of log entries from the networked system;

determining whether the second sequence is anomalous based on comparing the second sequence against the predicted log entries;

in response to determining that the second sequence is anomalous, determining whether the second sequence is noteworthy based on a function entropy of the first sequence and a sentiment polarity of the second sequence, wherein the function entropy is a measure of randomness in the loci entries, and wherein the sentiment polarity is a measure of sentiment of words found in the log entries; and

in response to determining that the second sequence is noteworthy, generating an anomaly report that includes the second sequence and a root cause.

16. The device of claim 15 , the operation further comprising:

training the machine learning model based on non-anomalous log entries for the networked system; and

maintaining a plurality of log templates for the log entries, wherein the plurality of log templates is updated during operation of the machine learning model to include new log templates observed in actual log entries from the networked system, further comprising:

adding a given log template to the plurality of log templates after the given log template is observed in the log entries a threshold number of times.

17. The device of claim 16 , wherein the machine learning model is an LSTM auto-encoder operating with an attention layer that is fitted with unlabeled log entries from the networked system, wherein the attention layer provides a vector representation based on outputs of the LSTM auto-encoder.

18. The device of claim 16 , wherein the predicted log entries are a set of log entries from the plurality of log templates with a highest probability based on the first sequence.

19. The device of claim 15 , wherein the second sequence is marked as not noteworthy when the function entropy is above an entropy threshold.

20. The device of claim 15 , wherein the second sequence is marked as not noteworthy when the sentiment polarity is above a positive sentiment threshold.