fix(component validation): make tests deterministic through absolute comparisons instead of bounds checks

[neuronull]

Changes

• Fixes a bug which was occasionally circumvented by comparing against a lower bound for some of the _bytes metrics- it seems the offered reasoning for the lower bound (inability to synchronize with the component's internal logic) was inaccurate.
  • The bug itself was that we were converting the EventData read from the yaml file, into a proper Event (and thus generating metadata for it), in two places- first when sending the event as input in the input runner, and again in the validator to construct the expected result. Because the event construction adds a timestamp, and because our determination of timestamp precision is based on whether there are trailing zeros that can be trimmed, this created a scenario where the actual event input to the component under test, and the event used to calculated expected metrics, had a timestamp that varied in it's number of bytes. This of course did not happen all of the time, which meant that the tests sometimes passed and sometimes failed.
  • The fix for this is to create the proper Event immediately after parsing the EventData from the yaml file (thereby only creating it once).
• Moves the calculation of expected metrics, out of the Validators and up to the input/output runners. This removes the need for a lot of logic that was previously being done in the Validators.
• Increases the static sleep we have when initializing the runner topology from 1s to 2s, and additionally increased the number of internal metric ticks to wait for before shutting down the telemetry from 2 ticks to 3. I found through experimentation that occasionally we would receive no events. The two attributes were increased independently and it was only through increasing both that the pass rate became 100%.
• Extracts duplicated code in the sources validator functions into helper functions.

Testing Done

The tests were run in a loop, stopping on the first failure.
This method was used to calibrate the waits we have on the topology as mentioned above.
Before the current settings, the maximum sequential passes was about 100.
With the current settings, the loop was manually stopped at about 1.2k iterations.

[netlify]

✅ Deploy Preview for vrl-playground canceled.

Name	Link
🔨 Latest commit	b7a7bd3
🔍 Latest deploy log	https://app.netlify.com/sites/vrl-playground/deploys/64b803ca2a71b90009a92873

[netlify]

✅ Deploy Preview for vector-project canceled.

Name	Link
🔨 Latest commit	b7a7bd3
🔍 Latest deploy log	https://app.netlify.com/sites/vector-project/deploys/64b803cae75dfb000885e005

[datadog-vectordotdev]

Datadog Report

Branch report: neuronull/draft_component_validation_better_validation
Commit report: caef34a

✅ vector: 0 Failed, 0 New Flaky, 1921 Passed, 0 Skipped, 1m 21.21s Wall Time

[tobz]

Just some general notes/ideas as I perused over this.

[tobz]

This is a lot cleaner overall, nice.

One random nit: we'll want to be careful that the codec emits all necessary bytes for the payload. Right now, as an example, for the HTTP resource, we're tracking the payload size because we only care about the body size... and the http_server source only tracks the body payload bytes, I guess. That's all fine and good, and matches how we're using the encoded size of the event here as the expected size.

However, you could imagine a scenario where maybe we track literally every byte that comes in over the HTTP connection -- request headers and all -- and then this approach would break down.

I can tell you that we definitely do that sort of approach for some components, such as the vector source, where we wrap the connection and count the literal bytes that flowed over the TCP socket. With this approach, and having a naive encoder implementation that simply encoded the event to its binary Protocol Buffers form... we'd end up with a byte size mismatch.

It might end up being easier at some point to move the collecting of these metrics to the resources themselves so that they can account for that, but I think this is a good step for right now because I do believe this logic will cover the majority of components.

[neuronull]

That's good to know ... I considered putting this in the resources themselves, but as I wasn't thinking of this concern, it seemed better at the time to put it where I did to simplify it and not require each resource to do it.

Will keep this in mind though going forward.

[neuronull]

Hmm. I started working on the sink side validators and ran into this with the http sink SentBytesTotal . Per the component spec

For HTTP-based protocols, the total number of bytes in the HTTP body, as represented by the Content-Length header.

I checked and indeed in the external resource it's got the correct byte count.

Unless I'm missing something, it seems like we're going to have to push this logic into the external resources now.

[tobz]

Ah yeah, not too surprising. 😬

[tobz]

LGTM, with one nit.

Obviously, I left the comment about potentially needing to move the metrics collection into the resources themselves to get full access to framing length, etc... but that could be a follow-up PR, really.

[neuronull]

Obviously, I left the comment about potentially needing to move the metrics collection into the resources themselves to get full access to framing length, etc... but that could be a follow-up PR, really.

👍

I started down that path in #17980 . If it becomes necessary, I create a separate issue to track this (if it doesn't get completely worked out in 17980).

[github-actions]

Regression Detector Results

Run ID: 932d8fb0-6f65-4d84-bbf8-4fa573260de5
Baseline: 7d0db6b
Comparison: 52a8036
Total vector CPUs: 7

Explanation

A regression test is an integrated performance test for vector in a repeatable rig, with varying configuration for vector. What follows is a statistical summary of a brief vector run for each configuration across SHAs given above. The goal of these tests are to determine quickly if vector performance is changed and to what degree by a pull request.

Because a target's optimization goal performance in each experiment will vary somewhat each time it is run, we can only estimate mean differences in optimization goal relative to the baseline target. We express these differences as a percentage change relative to the baseline target, denoted "Δ mean %". These estimates are made to a precision that balances accuracy and cost control. We represent this precision as a 90.00% confidence interval denoted "Δ mean % CI": there is a 90.00% chance that the true value of "Δ mean %" is in that interval.

We decide whether a change in performance is a "regression" -- a change worth investigating further -- if both of the following two criteria are true:

1. The estimated |Δ mean %| ≥ 5.00%. This criterion intends to answer the question "Does the estimated change in mean optimization goal performance have a meaningful impact on your customers?". We assume that when |Δ mean %| < 5.00%, the impact on your customers is not meaningful. We also assume that a performance change in optimization goal is worth investigating whether it is an increase or decrease, so long as the magnitude of the change is sufficiently large.

2. Zero is not in the 90.00% confidence interval "Δ mean % CI" about "Δ mean %". This statement is equivalent to saying that there is at least a 90.00% chance that the mean difference in optimization goal is not zero. This criterion intends to answer the question, "Is there a statistically significant difference in mean optimization goal performance?". It also means there is no more than a 10.00% chance this criterion reports a statistically significant difference when the true difference in mean optimization goal is zero -- a "false positive". We assume you are willing to accept a 10.00% chance of inaccurately detecting a change in performance when no true difference exists.

The table below, if present, lists those experiments that have experienced a statistically significant change in mean optimization goal performance between baseline and comparison SHAs with 90.00% confidence OR have been detected as newly erratic. Negative values of "Δ mean %" mean that baseline is faster, whereas positive values of "Δ mean %" mean that comparison is faster. Results that do not exhibit more than a ±5.00% change in their mean optimization goal are discarded. An experiment is erratic if its coefficient of variation is greater than 0.1. The abbreviated table will be omitted if no interesting change is observed.

No interesting changes in experiment optimization goals with confidence ≥ 90.00% and |Δ mean %| ≥ 5.00%.

Fine details of change detection per experiment.

experiment	goal	Δ mean %	Δ mean % CI	confidence
syslog_log2metric_humio_metrics	ingress throughput	+4.02	[+3.89, +4.15]	100.00%
otlp_http_to_blackhole	ingress throughput	+3.86	[+3.69, +4.04]	100.00%
datadog_agent_remap_blackhole_acks	ingress throughput	+2.73	[+2.63, +2.83]	100.00%
syslog_regex_logs2metric_ddmetrics	ingress throughput	+1.90	[+1.63, +2.16]	100.00%
http_text_to_http_json	ingress throughput	+1.27	[+1.18, +1.36]	100.00%
otlp_grpc_to_blackhole	ingress throughput	+1.06	[+0.94, +1.17]	100.00%
datadog_agent_remap_blackhole	ingress throughput	+0.98	[+0.87, +1.08]	100.00%
datadog_agent_remap_datadog_logs_acks	ingress throughput	+0.67	[+0.57, +0.77]	100.00%
syslog_log2metric_splunk_hec_metrics	ingress throughput	+0.49	[+0.37, +0.60]	100.00%
enterprise_http_to_http	ingress throughput	+0.06	[+0.02, +0.10]	95.68%
splunk_hec_indexer_ack_blackhole	ingress throughput	+0.01	[-0.03, +0.06]	33.35%
splunk_hec_to_splunk_hec_logs_noack	ingress throughput	+0.01	[-0.04, +0.05]	18.33%
fluent_elasticsearch	ingress throughput	+0.00	[-0.00, +0.00]	14.78%
http_to_http_json	ingress throughput	-0.00	[-0.04, +0.04]	9.85%
splunk_hec_to_splunk_hec_logs_acks	ingress throughput	-0.01	[-0.07, +0.06]	13.25%
socket_to_socket_blackhole	ingress throughput	-0.02	[-0.09, +0.04]	36.63%
http_to_http_noack	ingress throughput	-0.09	[-0.15, -0.03]	95.45%
datadog_agent_remap_datadog_logs	ingress throughput	-0.22	[-0.33, -0.11]	98.96%
syslog_humio_logs	ingress throughput	-1.00	[-1.11, -0.90]	100.00%
http_to_http_acks	ingress throughput	-1.67	[-2.90, -0.44]	91.85%
syslog_splunk_hec_logs	ingress throughput	-1.88	[-1.97, -1.80]	100.00%
file_to_blackhole	egress throughput	-2.41	[-6.02, +1.20]	60.78%
splunk_hec_route_s3	ingress throughput	-3.56	[-3.70, -3.42]	100.00%
syslog_loki	ingress throughput	-3.97	[-4.05, -3.89]	100.00%