An application is writing log events to Amazon CloudWatch Logs, and one field records application latency. The administrator needs to track the p90 (90th percentile) of that latency field over time. What should be done to collect that metric?

Creating a metric filter is the correct approach because it allows you to extract numerical values from log events and transform them into CloudWatch metrics. You can define a filter pattern to match log events containing the latency field and then specify the field to be extracted as the metric value. CloudWatch can then aggregate these extracted values and calculate statistics like p90 (90th percentile) over time. Contributor Insights is used for analyzing high-cardinality data and identifying top contributors, not for extracting specific percentiles from a numerical field. A subscription filter sends log events to other AWS services for processing, but it doesn't directly create CloudWatch metrics for percentile tracking. CloudWatch Application Insights is for monitoring application health and performance, but it doesn't provide the direct mechanism to create custom percentile metrics from arbitrary log fields like a metric filter does.