[ContentUnderstanding][P1] Python SDK default polling_interval (30s) is too high

[chienyuanchang]

[ContentUnderstanding][P1] Python SDK default polling_interval (30s) is too high

Language: Python
Notebook scenario: All scenarios using begin_analyze / begin_analyze_binary

Repro Steps

1. Create a ContentUnderstandingClient with all defaults (no polling_interval override)
2. Call begin_analyze_binary() with a small PDF (e.g., sample_invoice.pdf, ~100KB)
3. Measure E2E time from begin_analyze_binary() to .result()

Expected Result

E2E time should be ~7-15s, matching actual server processing time.

Actual Result

E2E time is ~33s due to default polling_interval=30 in ContentUnderstandingClientConfiguration.

The first GET poll doesn't happen until 30s after the POST, even though the operation typically completes in ~7-15s. This means the client sits idle for ~15-25s before discovering the result is ready.

Evidence from testing:

• polling_interval=1: E2E = ~7s (POST ~2.5s + 3-4 GET polls at ~1s intervals)
• polling_interval=30 (default): E2E = ~33s (POST ~2.5s + 30s wait + 1 GET poll)

Root Cause

ContentUnderstandingClientConfiguration sets self.polling_interval = kwargs.get("polling_interval", 30). This 30s default is too conservative for Content Understanding operations.

Suggestion

Consider lowering the default to 1-5 seconds, or implement exponential backoff starting from a short interval (e.g., 1s → 2s → 4s → 8s → ...).

The service does not return a Retry-After header, so the SDK purely relies on polling_interval.

[chienyuanchang]

Set to 3 seconds in GA SDK.