Extracting High-Level Metrics

In the previous step, we pre-processed the data to make it easier to analyze in SQL. Next, we want to generate valuable insights from it.

P95 Latency

P95 latency is a common statistical measure used for network performance analysis. It means that 95% of requests were served faster than this value, while 5% took longer than this.

Step 1: Define a Tumbling Window

To compute P95 latency over time, we first define a 10-second tumbling window on the spans view we defined previously.

-- continuation of the SQL program from the previous page

CREATE LOCAL VIEW spans_tumble_10s AS
SELECT * FROM TABLE(
	TUMBLE(
		TABLE spans,
		descriptor(eventtime),
		INTERVAL '10' SECOND
	)
);

This groups spans into non-overlapping 10-second time windows.

Step 2: Define a Rust UDF for P95 Calculation

Typically, the SQL function PERCENTILE can be used to calculate the P95 latency. However, Feldera SQL doesn't support it yet. This presents us an opportunity to use a Rust UDFs. This UDF will be a function that given an array of integers, returns the 95th percentile value from it.

SQL Definition for P95 Rust UDF

-- UDF to calculate p95 value given an integer array
CREATE FUNCTION p95(x BIGINT ARRAY NOT NULL) RETURNS BIGINT;

Rust Implementation of P95

// udf.rs

pub fn p95(x: Vec<Option<i64>>) -> Result<Option<i64>, Box<dyn std::error::Error>> {
	let mut x: Vec<i64> = x.into_iter().filter_map(|x| x).collect();

	if x.is_empty() {
		return Ok(None);
	}

	x.sort_unstable();

	let rank = 0.95 * x.len() as f64 - 1.0;
	let lower = rank.floor() as usize;
	let upper = rank.ceil() as usize;

	if lower == upper {
		Ok(Some(x[lower]))
	} else {
		let weight = rank - lower as f64;
		Ok(Some((x[lower] as f64 * (1.0 - weight) + x[upper] as f64 * weight) as i64))
	}
}

This function:

1. Filters out null values and sorts the values.
2. Finds the 95th percentile index.
3. Interpolates and returns the P95 value.

Step 3: Compute P95 Latency per 10 Second Window

-- Calculate the p95 latency in milliseconds
CREATE MATERIALIZED VIEW p95_latency AS
SELECT
	p95(ARRAY_AGG(elapsedtimemillis)) AS latencyms,
	window_start AS 'time'
FROM spans_tumble_10s
WHERE
	parentspanid = '' -- Only consider top-level requests
GROUP BY
	window_start;

This filters only the top level traces and aggregates elapsedTimeMillis values into an array and calls the p95 function.

Throughput

Throughput measures the number of top-level requests processed per second. In this demo however, we calculate throughput on 10 second time buckets instead of per second.

CREATE MATERIALIZED VIEW throughput AS
SELECT
	COUNT(*) AS throughput,
	window_start AS 'time'
FROM
	spans_tumble_10s
WHERE
	parentspanid = ''
GROUP BY
	window_start;

Operation Execution Time

Operation execution time measures how long each operation actually took, excluding time spent in child spans.

CREATE MATERIALIZED VIEW operation_execution_time AS
SELECT
    s.name,
    SUM(
        s.elapsedtimemillis -
        COALESCE(
            SELECT
                SUM(elapsedtimemillis)
                FROM spans k
                WHERE k.traceid = s.traceid
                AND k.parentspanid = s.spanid,
            0
        )
    ) AS elapsed
FROM spans s
GROUP BY s.name;

Explanation: 4. s.elapsedTimeMillis: Total execution time of the operation. 5. Subtracts time spent in child spans to compute actual execution time. 6. Groups results by operation name s.name.