Batch vs real-time inference in NLP - Metrics Comparison

Metrics & Evaluation - Batch vs real-time inference

Which metric matters for Batch vs Real-time inference and WHY

For batch inference, throughput and latency matter because you process many inputs at once and want to finish quickly.

For real-time inference, latency is the most important metric because users expect fast responses.

Accuracy, precision, and recall still matter for model quality, but performance metrics like latency and throughput decide if the system meets user needs.

Confusion matrix example (for classification quality)

      | Predicted Positive | Predicted Negative |
      |--------------------|--------------------|
      | True Positive (TP)  | False Positive (FP) |
      | False Negative (FN) | True Negative (TN)  |

      Example numbers for 1000 samples:
      TP=400, FP=100, FN=50, TN=450

      Total = TP + FP + FN + TN = 1000
    

This confusion matrix is the same for batch or real-time inference since it measures model correctness.

Tradeoff: Precision vs Recall in Batch vs Real-time

In batch inference, you can afford to tune for higher precision because you have time to review or reprocess results.

In real-time inference, you might prioritize recall to catch as many important cases as possible quickly, even if some false alarms happen.

Example: A spam filter in real-time should catch most spam (high recall) to protect users immediately.

What "good" vs "bad" metric values look like

Batch inference: Good throughput (e.g., 1000 samples/sec), acceptable latency (e.g., minutes), and high accuracy (e.g., 95%).

Real-time inference: Low latency (e.g., under 100 ms per request), stable throughput (e.g., 10 requests/sec), and high recall (e.g., 90%) for critical cases.

Bad values mean slow responses in real-time or very low accuracy in both modes.

Common pitfalls in metrics for Batch vs Real-time inference

• Ignoring latency in real-time systems leads to poor user experience.
• Measuring only accuracy without considering latency or throughput.
• Data leakage causing inflated accuracy in batch evaluation but poor real-time results.
• Overfitting to batch data that does not represent real-time input distribution.

Self-check question

Your real-time model has 98% accuracy but only 12% recall on fraud cases. Is it good for production? Why or why not?

Answer: No, it is not good. Even though accuracy is high, the model misses most fraud cases (low recall). In fraud detection, missing fraud is very costly, so recall is more important.

Key Result

Latency is key for real-time inference; throughput and accuracy matter more for batch inference.

Practice

1. What is the main difference between batch inference and real-time inference in NLP?

easy

A. Batch inference requires internet connection, real-time inference does not.

B. Batch inference is slower than real-time inference because it uses outdated models.

C. Real-time inference processes data only at night, batch inference runs during the day.

D. Batch inference processes many inputs together, while real-time inference processes inputs one by one quickly.

Solution

1. Step 1: Understand batch inference

   Batch inference means processing many inputs together in one go, which is efficient for large data.

2. Step 2: Understand real-time inference

   Real-time inference means processing each input immediately to give instant results.

3. Final Answer:

   Batch inference processes many inputs together, while real-time inference processes inputs one by one quickly. -> Option D

4. Quick Check:

   Batch = many inputs, Real-time = instant input [OK]

Hint: Batch = many at once, real-time = one fast [OK]

Common Mistakes:

• Confusing batch with outdated models
• Thinking real-time only runs at specific times
• Mixing internet requirements

2. Which code snippet correctly represents a batch inference call for an NLP model?

easy

A. model.load('batch')

B. model.predict('text1')

C. model.predict(['text1', 'text2', 'text3'])

D. model.train(['text1', 'text2'])

Solution

1. Step 1: Identify batch input format

   Batch inference requires passing multiple inputs together, usually as a list or array.

2. Step 2: Check code options

   model.predict(['text1', 'text2', 'text3']) passes a list of texts to predict, which is correct for batch inference.

3. Final Answer:

   model.predict(['text1', 'text2', 'text3']) -> Option C

4. Quick Check:

   Batch input = list of texts [OK]

Hint: Batch inference uses list input for prediction [OK]

Common Mistakes:

• Passing single string instead of list
• Confusing training with inference
• Using unrelated method like load

3. Given the code below, what will be the output type of results?

texts = ['hello', 'world']
results = model.predict(texts)

Assuming model.predict returns predictions for each input.

medium

A. A list of predictions, one for each input text

B. A single prediction combining all texts

C. An error because input is a list

D. A dictionary with input texts as keys

Solution

1. Step 1: Understand input to model.predict

   The input is a list of texts, so the model will process each text separately.

2. Step 2: Understand output type for batch input

   For batch input, the output is usually a list of predictions, matching the input size.

3. Final Answer:

   A list of predictions, one for each input text -> Option A

4. Quick Check:

   Batch input gives list output [OK]

Hint: Batch input returns list output matching inputs [OK]

Common Mistakes:

• Expecting single combined prediction
• Thinking list input causes error
• Assuming output is a dictionary

4. Identify the error in this real-time inference code snippet:

input_text = ['Hello world']
prediction = model.predict(input_text)

Assuming model.predict expects a single string for real-time inference.

medium

A. Input should be a string, not a list

B. model.predict cannot process text

C. Missing batch size parameter

D. Prediction variable name is invalid

Solution

1. Step 1: Check input type for real-time inference

   Real-time inference expects a single input string, not a list.

2. Step 2: Identify mismatch in code

   The code passes a list with one string, causing a type mismatch error.

3. Final Answer:

   Input should be a string, not a list -> Option A

4. Quick Check:

   Real-time input = string only [OK]

Hint: Real-time input must be a single string [OK]

Common Mistakes:

• Passing list instead of string
• Assuming batch size needed for real-time
• Thinking variable name causes error

5. You have a large dataset of 10,000 sentences to classify using an NLP model. You want to minimize total processing time but can wait a few minutes for results. Which inference method should you choose and why?

hard

A. Neither, you should retrain the model first.

B. Batch inference, because processing many inputs together is more efficient for large data.

C. Real-time inference, because it processes each sentence instantly.

D. Real-time inference, because it uses less memory.

Solution

1. Step 1: Analyze dataset size and time constraints

   With 10,000 sentences and willingness to wait minutes, efficiency matters more than instant results.

2. Step 2: Choose inference method based on efficiency

   Batch inference processes many inputs together, reducing overhead and total time.

3. Final Answer:

   Batch inference, because processing many inputs together is more efficient for large data. -> Option B

4. Quick Check:

   Large data + wait time = batch inference [OK]

Hint: Large data with wait time? Use batch inference [OK]

Common Mistakes:

• Choosing real-time for large batch
• Thinking retraining is needed
• Assuming real-time uses less memory always