Audit trails for model decisions in MLOps - Step-by-Step Execution

Process Flow - Audit trails for model decisions

Model receives input data

↓

Model makes decision

↓

Log decision details

↓

Store logs in audit trail system

↓

Audit trail available for review

↓

Use audit trail for debugging, compliance, or improvement

This flow shows how a model decision is logged step-by-step to create an audit trail for later review and compliance.

Execution Sample

MLOps

input_data = {'age': 30, 'income': 50000}
decision = model.predict(input_data)
audit_log = {
  'input': input_data,
  'decision': decision,
  'timestamp': current_time()
}
store_audit_log(audit_log)

This code logs the input, model decision, and timestamp to an audit trail system.

Process Table

Step	Action	Data Captured	Storage Location	Result
1	Receive input data	{'age': 30, 'income': 50000}	In memory	Input ready for prediction
2	Model makes decision	Prediction result (e.g., 'approve')	In memory	Decision generated
3	Create audit log entry	{input, decision, timestamp}	In memory object	Audit log prepared
4	Store audit log	Audit log object	Audit trail database or file	Audit log saved
5	Audit trail available	Stored logs	Audit trail system	Logs ready for review

💡 Audit trail is complete after storing the log for the model decision

Status Tracker

Variable	Start	After Step 1	After Step 2	After Step 3	Final
input_data	None	{'age': 30, 'income': 50000}	{'age': 30, 'income': 50000}	{'age': 30, 'income': 50000}	{'age': 30, 'income': 50000}
decision	None	None	'approve'	'approve'	'approve'
audit_log	None	None	None	{'input': {'age': 30, 'income': 50000}, 'decision': 'approve', 'timestamp': '2024-06-01T12:00:00Z'}	{'input': {'age': 30, 'income': 50000}, 'decision': 'approve', 'timestamp': '2024-06-01T12:00:00Z'}

Key Moments - 3 Insights

Why do we log both input data and model decision in the audit trail?

Logging both input and decision ensures we can trace exactly what data led to each decision, as shown in execution_table rows 1 and 2.

What happens if the audit log is not stored properly after creation?

If the audit log is not saved (execution_table step 4), the audit trail is incomplete and cannot be reviewed later, losing traceability.

Why is a timestamp included in the audit log?

Timestamp records when the decision was made, helping order events and support compliance, as seen in the audit_log variable at step 3.

Visual Quiz - 3 Questions

Look at the execution_table at step 2. What data is captured?

A Input data only

B Prediction result only

C Input data and prediction result

D Timestamp only

💡 Hint

Check the 'Data Captured' column at step 2 in the execution_table.

At which step is the audit log saved to permanent storage?

A Step 4

B Step 3

C Step 1

D Step 5

💡 Hint

Look for the step where 'Store audit log' action happens in the execution_table.

If the input_data changes, which variable_tracker column will show the difference?

A After Step 2

B After Step 3

C After Step 1

D Final

💡 Hint

Check when input_data is first assigned in variable_tracker.

Concept Snapshot

Audit trails log model inputs, decisions, and timestamps.
Store logs in a secure system for review.
Helps with debugging, compliance, and trust.
Always capture input, output, and time.
Ensure logs are saved reliably.

Full Transcript

Audit trails for model decisions capture the input data, the model's decision, and the time the decision was made. This information is stored in a secure audit trail system. The process starts when the model receives input data, makes a decision, then creates a log entry with all details. This log is saved for later review to help with debugging, compliance, and improving the model. Key variables like input_data, decision, and audit_log change step-by-step as the process runs. Understanding when and what is logged helps ensure traceability and trust in automated decisions.

Practice

1. What is the main purpose of audit trails in machine learning model decisions?

easy

A. To encrypt the model data for security

B. To speed up the model training process

C. To reduce the size of the model

D. To record inputs, outputs, and context for each model decision

Solution

1. Step 1: Understand audit trail purpose

   Audit trails are used to keep a record of what data was input, what output was produced, and the context around the decision.

2. Step 2: Compare options

   Only To record inputs, outputs, and context for each model decision describes this purpose correctly. Other options describe unrelated tasks.

3. Final Answer:

   To record inputs, outputs, and context for each model decision -> Option D

4. Quick Check:

   Audit trails = record inputs and outputs [OK]

Hint: Audit trails track what goes in and out of models [OK]

Common Mistakes:

• Confusing audit trails with model optimization
• Thinking audit trails speed up training
• Believing audit trails encrypt data

2. Which of the following is the correct way to log a model decision with timestamp in Python?

easy

A. log_entry = f"{datetime.now()} - Input: {input_data}, Output: {output}"

B. log_entry = datetime.now() + input_data + output

C. log_entry = "Input: input_data, Output: output"

D. log_entry = f"Input: {input_data} Output: {output}"

Solution

1. Step 1: Check correct string formatting with timestamp

   log_entry = f"{datetime.now()} - Input: {input_data}, Output: {output}" uses f-string with datetime.now() to include timestamp and variables properly.

2. Step 2: Identify errors in other options

   log_entry = datetime.now() + input_data + output tries to add incompatible types, causing error. Options C and D miss timestamp or variable interpolation.

3. Final Answer:

   log_entry = f"{datetime.now()} - Input: {input_data}, Output: {output}" -> Option A

4. Quick Check:

   Use f-string with datetime.now() for logging [OK]

Hint: Use f-strings and datetime.now() for timestamped logs [OK]

Common Mistakes:

• Concatenating incompatible types without conversion
• Forgetting to include timestamp
• Not using variable interpolation in strings

3. Given the following Python code snippet for logging model decisions, what will be the output?

from datetime import datetime
input_data = {'age': 30}
output = 'approved'
log_entry = f"{datetime(2024, 6, 1, 12, 0)} - Input: {input_data}, Output: {output}"
print(log_entry)

medium

A. 2024/06/01 12:00 - Input: {'age': 30}, Output: approved

B. datetime.datetime(2024, 6, 1, 12, 0) - Input: {'age': 30}, Output: approved

C. 2024-06-01 12:00:00 - Input: {'age': 30}, Output: approved

D. Error: datetime object cannot be formatted in f-string

Solution

1. Step 1: Understand datetime object formatting in f-string

   Using datetime(2024, 6, 1, 12, 0) in f-string calls its __str__ method, which outputs '2024-06-01 12:00:00'.

2. Step 2: Combine string parts

   The rest of the string includes input_data and output as expected, so the full string prints correctly.

3. Final Answer:

   2024-06-01 12:00:00 - Input: {'age': 30}, Output: approved -> Option C

4. Quick Check:

   Datetime __str__ = 'YYYY-MM-DD HH:MM:SS' [OK]

Hint: Datetime prints as 'YYYY-MM-DD HH:MM:SS' in f-strings [OK]

Common Mistakes:

• Expecting datetime object to print as constructor call
• Confusing date formats
• Thinking f-string cannot handle datetime objects

4. You have this code snippet to log model decisions but it raises an error:

log_entry = f"{datetime.now()} - Input: {input_data}, Output: {output}"

What is the most likely cause of the error?

medium

A. datetime module is not imported

B. input_data is not defined

C. f-string syntax is incorrect

D. output variable is a number, not a string

Solution

1. Step 1: Check for datetime usage

   Using datetime.now() requires importing datetime module or class. If missing, NameError occurs.

2. Step 2: Verify other variables and syntax

   input_data and output can be any type; f-string handles them. Syntax is correct.

3. Final Answer:

   datetime module is not imported -> Option A

4. Quick Check:

   Missing import datetime causes NameError [OK]

Hint: Always import datetime before using datetime.now() [OK]

Common Mistakes:

• Assuming variables cause error without checking imports
• Thinking f-string syntax is wrong
• Believing numbers cause f-string errors

5. You want to create an audit trail that records model version, input data, output, and timestamp in JSON format for each decision. Which Python code snippet correctly creates this audit trail entry?

hard

A. import json, datetime
audit_entry = json.dumps({"model_version": "v1.2", "input": input_data, "output": output, "timestamp": datetime.datetime.now.isoformat()})

B. import json, datetime
audit_entry = json.dumps({"model_version": "v1.2", "input": input_data, "output": output, "timestamp": datetime.now().isoformat()})

C. import json, datetime
audit_entry = json.dumps({"model_version": "v1.2", "input": input_data, "output": output, "timestamp": datetime.now().str()})

D. import json, datetime
audit_entry = json.dumps({"model_version": "v1.2", "input": input_data, "output": output, "timestamp": datetime.now()})

Solution

1. Step 1: Check correct import and datetime usage

   import json, datetime audit_entry = json.dumps({"model_version": "v1.2", "input": input_data, "output": output, "timestamp": datetime.now().isoformat()}) correctly imports datetime and uses datetime.now().isoformat() to get a string timestamp.

2. Step 2: Validate JSON serialization

   datetime.now() returns a datetime object which is not JSON serializable directly, so isoformat() converts it to string. import json, datetime audit_entry = json.dumps({"model_version": "v1.2", "input": input_data, "output": output, "timestamp": datetime.now()}) fails here.

3. Step 3: Check other options

   import json, datetime audit_entry = json.dumps({"model_version": "v1.2", "input": input_data, "output": output, "timestamp": datetime.datetime.now.isoformat()}) tries to call isoformat on the now method object (missing () after now), causing AttributeError. import json, datetime audit_entry = json.dumps({"model_version": "v1.2", "input": input_data, "output": output, "timestamp": datetime.now().str()}) tries to call .str() on datetime object, causing AttributeError.

4. Final Answer:

   import json, datetime audit_entry = json.dumps({"model_version": "v1.2", "input": input_data, "output": output, "timestamp": datetime.now().isoformat()}) -> Option B

5. Quick Check:

   Use datetime.now().isoformat() for JSON timestamp [OK]

Hint: Use datetime.now().isoformat() for JSON-friendly timestamps [OK]

Common Mistakes:

• Missing () after now() leading to method object error
• Trying to serialize datetime object directly
• Using non-existent .str() method on datetime