[DRAFT] fix(bigtable): standardize client side metrics

[daniel-sanche]

Made some adjustments to the client side metrics system after comparing against node and java implemenations (with gemini)

The diffs for each commit can be looked it in isolation

1 (0b3600d): throttling_latencies is now recorded once per operation

• java records grpc time as part of throttling latencies. Python just tracks flow-control time
• flow control happens once at the start of an operation: once the mutations have made it past the flow control gate, retries don't have to wait again, they have space reserved until they reach a terminal state
• Python was previously recording throttling_latencieson each attempt, but recording the same per-operation value on each attempt. This would give misleading data
• Now, we only record it once per operation

2 (b44895d): removing backoff time from application blocking latencies

• previously, python was including time spent backing off between attempts as part of the application_latencies metric
• other languages do not include this, so we can drop it from Python as well

[gemini-code-assist]

Code Review

This pull request refactors how throttling and application latencies are recorded in OpenTelemetry. Specifically, throttling latencies are now recorded on operation completion rather than attempt completion, and only for bulk mutate rows operations. Additionally, application latencies no longer include backoff time. The review feedback highlights a potential TypeError if flow_throttling_time_ns is None during the operation completion check, recommending a defensive check and an accompanying unit test to handle this case safely.

[gemini-code-assist]

If op.flow_throttling_time_ns is None, evaluating op.flow_throttling_time_ns > 0 will raise a TypeError at runtime. Adding a defensive check to ensure it is not None before comparison prevents potential crashes.

Suggested change

	if (
	op.op_type == OperationType.BULK_MUTATE_ROWS
	and op.flow_throttling_time_ns > 0
	):
	self.otel.throttling_latencies.record(
	op.flow_throttling_time_ns / NS_TO_MS, labels
	)
	if (
	op.op_type == OperationType.BULK_MUTATE_ROWS
	and op.flow_throttling_time_ns is not None
	and op.flow_throttling_time_ns > 0
	):
	self.otel.throttling_latencies.record(
	op.flow_throttling_time_ns / NS_TO_MS, labels
	)

[gemini-code-assist]

Add a test case where flow_throttling_ns is None to verify that the handler safely handles None values without raising a TypeError.

    @pytest.mark.parametrize(
        "op_type,flow_throttling_ns,should_record",
        [
            (OperationType.BULK_MUTATE_ROWS, 54321, True),
            (OperationType.BULK_MUTATE_ROWS, 0, False),
            (OperationType.BULK_MUTATE_ROWS, None, False),
            (OperationType.READ_ROWS, 54321, False),
        ],
    )