Machine learning anomaly detection in Elasticsearch - Step-by-Step Execution

Concept Flow - Machine learning anomaly detection

Input Data

↓

Feature Extraction

↓

Train ML Model

↓

Detect Anomalies

↓

Output Anomaly Scores

↓

Alert or Visualize

Data flows from input through feature extraction, model training, anomaly detection, and finally outputs anomaly scores for alerts or visualization.

Execution Sample

Elasticsearch

POST _ml/anomaly_detectors/my_detector/_evaluate
{
  "data": [
    {"timestamp": "2024-06-01T00:00:00Z", "value": 10},
    {"timestamp": "2024-06-01T01:00:00Z", "value": 1000}
  ]
}

This request sends sample data to an Elasticsearch ML anomaly detector to evaluate anomaly scores.

Execution Table

Step	Input Data Point	Feature Extracted	Model Score	Anomaly Score	Action
1	{"timestamp": "2024-06-01T00:00:00Z", "value": 10}	value=10	Normal	Low (0.1)	No alert
2	{"timestamp": "2024-06-01T01:00:00Z", "value": 1000}	value=1000	Anomalous	High (0.95)	Alert generated
3	End of data	-	-	-	Stop evaluation

💡 All data points processed, anomaly scores assigned, evaluation stops.

Variable Tracker

Variable	Start	After 1	After 2	Final
value	N/A	10	1000	1000
anomaly_score	N/A	0.1	0.95	0.95
alert	False	False	True	True

Key Moments - 2 Insights

Why does the anomaly score jump from 0.1 to 0.95 between the two data points?

Because the second value (1000) is very different from normal values seen before, the model flags it as anomalous, shown in execution_table row 2.

What does a low anomaly score mean in this context?

A low anomaly score means the data point fits the normal pattern learned by the model, so no alert is triggered, as seen in execution_table row 1.

Visual Quiz - 3 Questions

Look at the execution table, what is the anomaly score for the first data point?

A 1.0

B 0.95

C 0.1

D 0.5

💡 Hint

Check the 'Anomaly Score' column in row 1 of the execution_table.

At which step does the model generate an alert?

A Step 1

B Step 2

C Step 3

D No alert generated

💡 Hint

Look at the 'Action' column in the execution_table for when alert is True.

If the second data point had a value of 15 instead of 1000, how would the anomaly score likely change?

A It would stay low, similar to 0.1

B It would be very high, close to 0.95

C It would be exactly 0.5

D It would cause an error

💡 Hint

Refer to variable_tracker for how anomaly_score changes with value differences.

Concept Snapshot

Machine learning anomaly detection in Elasticsearch:
- Input data is processed to extract features.
- ML model learns normal patterns.
- Each data point gets an anomaly score.
- High scores indicate unusual data.
- Alerts or visualizations help monitor anomalies.

Full Transcript

Machine learning anomaly detection in Elasticsearch works by taking input data and extracting features like numeric values. The ML model is trained to understand what normal data looks like. When new data points arrive, the model scores them based on how unusual they are. Low anomaly scores mean the data fits normal patterns, while high scores mean the data is unusual and may indicate a problem. Alerts can be generated for high anomaly scores to notify users. This process helps monitor data streams for unexpected changes or errors.

Practice

1. What is the main purpose of machine learning anomaly detection in Elasticsearch?

easy

A. To automatically find unusual patterns in data

B. To store large amounts of data efficiently

C. To create visual dashboards for data

D. To backup Elasticsearch clusters

Solution

1. Step 1: Understand anomaly detection goal

   Machine learning anomaly detection is designed to find unusual or unexpected patterns in data automatically.

2. Step 2: Compare options with purpose

   Options B, C, and D describe other Elasticsearch features, not anomaly detection.

3. Final Answer:

   To automatically find unusual patterns in data -> Option A

4. Quick Check:

   Purpose of anomaly detection = find unusual patterns [OK]

Hint: Anomaly detection finds unusual data automatically [OK]

Common Mistakes:

• Confusing anomaly detection with data storage
• Thinking anomaly detection creates dashboards
• Mixing anomaly detection with backup tasks

2. Which Elasticsearch API call starts the anomaly detection process by feeding data to the job?

easy

A. POST _ml/anomaly_detectors/<job_id>/_start_datafeed

B. GET _ml/anomaly_detectors/<job_id>/results

C. PUT _ml/anomaly_detectors/<job_id>

D. DELETE _ml/anomaly_detectors/<job_id>

Solution

1. Step 1: Identify datafeed start API

   The API to start feeding data to an anomaly detection job is POST _ml/anomaly_detectors/<job_id>/_start_datafeed.

2. Step 2: Eliminate other options

   GET retrieves results, PUT creates or updates jobs, DELETE removes jobs.

3. Final Answer:

   POST _ml/anomaly_detectors/<job_id>/_start_datafeed -> Option A

4. Quick Check:

   Start datafeed = POST _start_datafeed [OK]

Hint: Start datafeed uses POST with _start_datafeed endpoint [OK]

Common Mistakes:

• Using GET instead of POST to start datafeed
• Confusing job creation with starting datafeed
• Deleting job instead of starting datafeed

3. Given this Elasticsearch ML job result snippet:

{"job_id":"sales_anomaly","results":[{"timestamp":1680000000000,"anomaly_score":75},{"timestamp":1680003600000,"anomaly_score":5}]}

Which timestamp shows a likely anomaly?

medium

A. Neither timestamp

B. 1680003600000

C. Both timestamps

D. 1680000000000

Solution

1. Step 1: Understand anomaly score meaning

   Higher anomaly scores indicate more unusual data points. A score of 75 is high, 5 is low.

2. Step 2: Identify timestamp with high score

   The timestamp 1680000000000 has anomaly_score 75, indicating a likely anomaly.

3. Final Answer:

   1680000000000 -> Option D

4. Quick Check:

   High anomaly score = likely anomaly [OK]

Hint: Higher anomaly_score means more likely anomaly [OK]

Common Mistakes:

• Choosing low anomaly score as anomaly
• Selecting both timestamps without checking scores
• Ignoring anomaly_score values

4. You created an anomaly detection job but see no results after starting the datafeed. What is a likely cause?

medium

A. The job was deleted before starting

B. The Elasticsearch cluster is offline

C. The datafeed is not running or has stopped

D. The anomaly scores are all zero

Solution

1. Step 1: Check datafeed status

   If no results appear, the datafeed may not be running or has stopped feeding data to the job.

2. Step 2: Evaluate other options

   Job deletion would prevent starting datafeed; cluster offline causes broader failures; zero scores still produce results.

3. Final Answer:

   The datafeed is not running or has stopped -> Option C

4. Quick Check:

   No results usually mean datafeed stopped [OK]

Hint: No results? Check if datafeed is running [OK]

Common Mistakes:

• Assuming zero scores mean no results
• Ignoring datafeed status
• Blaming cluster offline without checking datafeed

5. You want to detect unusual spikes in website traffic using Elasticsearch ML anomaly detection. Which steps should you follow to set this up correctly?

hard

A. Backup traffic data, create index pattern, then visualize spikes

B. Create a job with traffic data, start datafeed, then analyze anomaly results

C. Create a dashboard, upload traffic logs, then run anomaly detection manually

D. Delete old data, create job without datafeed, then check results

Solution

1. Step 1: Create ML job with traffic data

   Define an anomaly detection job using the website traffic data to analyze patterns.

2. Step 2: Start the datafeed to feed data into the job

   Start the datafeed so the job can process incoming traffic data continuously.

3. Step 3: Analyze the anomaly detection results

   Review the results to identify unusual spikes or anomalies in traffic.

4. Final Answer:

   Create a job with traffic data, start datafeed, then analyze anomaly results -> Option B

5. Quick Check:

   Job + datafeed + analyze = correct setup [OK]

Hint: Job creation + datafeed start + result check = setup [OK]

Common Mistakes:

• Skipping datafeed start step
• Confusing dashboards with anomaly detection setup
• Deleting data before analysis