Add blog post: Moving local flag evaluation from Django to Rust by jina-yoon · Pull Request #16882 · PostHog/posthog.com

jina-yoon · 2026-05-19T00:51:13Z

Summary

Adds Patricio's blog post on migrating the feature flags local evaluation endpoint from Django to Rust
Adds Patricio Tarantino to authors.json as a Product Engineer
Post lives at /blog/feature-flags-definitions-rust

To do before merging

Replace placeholder featuredImage URL in frontmatter with the actual Cloudinary image
Confirm the /blog/feature-flags-service link at the bottom is correct (no matching post found in the repo — may need a redirect or URL update)

🤖 Generated with Claude Code

…ation Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

github-actions · 2026-05-19T00:51:26Z

Deploy preview

Status	Details	Updated (UTC)
🟢 Ready	View preview	May 19, 2026 02:06AM

github-actions · 2026-05-19T00:52:14Z

Vale prose linter → found 2 errors, 19 warnings, 0 suggestions in your markdown

Full report → Copy the linter results into an LLM to batch-fix issues.

Linter being weird? Update the rules!

contents/blog/even-faster-more-reliable-flags.md — 2 errors, 8 warnings, 0 suggestions

Line	Severity	Message	Rule
2:20	warning	Capitalize 'Feature Flags' for PostHog's product. Use 'feature flags' for the general industry concept.	`PostHogBase.ProductNames`
18:27	warning	Capitalize 'Feature Flags' for PostHog's product. Use 'feature flags' for the general industry concept.	`PostHogBase.ProductNames`
24:2	warning	Capitalize 'Feature Flags' for PostHog's product. Use 'Feature flags' for the general industry concept.	`PostHogBase.ProductNames`
26:95	warning	Capitalize 'Feature Flags' for PostHog's product. Use 'feature flags' for the general industry concept.	`PostHogBase.ProductNames`
52:15	warning	'axum' is a possible misspelling.	`PostHogBase.Spelling`
54:46	warning	'axum' is a possible misspelling.	`PostHogBase.Spelling`
54:126	error	Hi, Andy here... use an en dash ( – ) with spaces. On Mac, holding down the Option and hyphen key will give you an en dash.	`PostHogBase.EnDash`
95:39	error	Hi, Andy here... use an en dash ( – ) with spaces. On Mac, holding down the Option and hyphen key will give you an en dash.	`PostHogBase.EnDash`
107:47	warning	'achilles' is a possible misspelling.	`PostHogBase.Spelling`
115:85	warning	Capitalize 'Feature Flags' for PostHog's product. Use 'Feature flags' for the general industry concept.	`PostHogBase.ProductNames`

contents/blog/local-flag-evals-rust.md — 0 errors, 11 warnings, 0 suggestions

Line	Severity	Message	Rule
25:346	warning	Capitalize 'Feature Flags' for PostHog's product. Use 'feature flags' for the general industry concept.	`PostHogBase.ProductNames`
31:174	warning	Capitalize 'Feature Flags' for PostHog's product. Use 'feature flags' for the general industry concept.	`PostHogBase.ProductNames`
37:161	warning	'mixin' is a possible misspelling.	`PostHogBase.Spelling`
39:3	warning	'ETags' is a possible misspelling.	`PostHogBase.Spelling`
39:170	warning	'ETags' is a possible misspelling.	`PostHogBase.Spelling`
41:79	warning	'allowlist' is a possible misspelling.	`PostHogBase.Spelling`
41:122	warning	'reimplemented' is a possible misspelling.	`PostHogBase.Spelling`
45:16	warning	'bursty' is a possible misspelling.	`PostHogBase.Spelling`
70:3	warning	'PGBouncer' is a possible misspelling.	`PostHogBase.Spelling`
81:140	warning	'Karpenter' is a possible misspelling.	`PostHogBase.Spelling`
85:21	warning	'Tokio' is a possible misspelling.	`PostHogBase.Spelling`

jina-yoon · 2026-05-19T01:06:57Z

@@ -0,0 +1,85 @@
+---
+title: 'Moving local flag evaluation from Django to Rust: 24x less CPU, 56x less memory'
+date: 2026-05-18


Suggested change

date: 2026-05-18

date: 2026-05-19

jina-yoon · 2026-05-19T01:10:07Z

+  - Feature flags
+seo: {
+  metaTitle: "Moving local flag evaluation from Django to Rust: 24x less CPU, 56x less memory",
+  metaDescription: "How we moved PostHog's feature flags local evaluation endpoint from Django to Rust, dropping CPU usage by 24x and memory by 56x."


Suggested change

metaDescription: "How we moved PostHog's feature flags local evaluation endpoint from Django to Rust, dropping CPU usage by 24x and memory by 56x."

metaDescription: "How we moved PostHog's feature flags local evaluation endpoint from Django to Rust and dropped CPU usage by 24x and memory by 56x."

jina-yoon · 2026-05-19T01:12:22Z

+
+---
+
+I reloaded Grafana three times before I trusted what I was looking at, because the gap between the old numbers and the new ones was bigger than I was expecting. p50 latency had gone from 40ms down to 4ms, CPU usage was sitting at a small fraction of where it used to be, and memory was barely registering at all. We had just finished moving the feature flags local evaluation endpoint from Django to Rust, and despite knowing that it was going to be better – that was the whole reason for this migration – I still wasn't ready for how the comparison ended up looking.


Suggested change

I reloaded Grafana three times before I trusted what I was looking at, because the gap between the old numbers and the new ones was bigger than I was expecting. p50 latency had gone from 40ms down to 4ms, CPU usage was sitting at a small fraction of where it used to be, and memory was barely registering at all. We had just finished moving the feature flags local evaluation endpoint from Django to Rust, and despite knowing that it was going to be better – that was the whole reason for this migration – I still wasn't ready for how the comparison ended up looking.

I reloaded Grafana three times before I could trust what I was looking at. I knew it was going to be better – that was the whole reason for this migration – but I still wasn't ready for how big of a difference it would make. p50 latency had gone from 40ms down to 4ms, CPU usage was sitting at a small fraction of where it used to be, and memory was barely registering at all.

Here's how we moved our local flag evaluation from Django to Rust.

jina-yoon · 2026-05-19T01:13:46Z

+| PGBouncer sidecars | Yes | No |
+| Dedicated node pool | Yes (memory-optimized) | No (shared pool) |
+
+Latency improved too. These numbers are measured at the Envoy layer (the time between Envoy receiving the request and getting a response back from the pod), so they don't include client-to-Envoy network time. A customer's SDK sees this plus its own network round-trip on top. Same Prometheus metric (`envoy_cluster_upstream_rq_time`) for both services, so it's apples to apples.


Suggested change

Latency improved too. These numbers are measured at the Envoy layer (the time between Envoy receiving the request and getting a response back from the pod), so they don't include client-to-Envoy network time. A customer's SDK sees this plus its own network round-trip on top. Same Prometheus metric (`envoy_cluster_upstream_rq_time`) for both services, so it's apples to apples.

Latency improved, too. These numbers are measured at the Envoy layer (the time between Envoy receiving the request and getting a response back from the pod), so they don't include client-to-Envoy network time. A customer's SDK sees this plus its own network round-trip on top. Since it's the same Prometheus metric (`envoy_cluster_upstream_rq_time`) for both services, it's apples to apples.

jina-yoon · 2026-05-19T01:15:28Z

+
+I reloaded Grafana three times before I trusted what I was looking at, because the gap between the old numbers and the new ones was bigger than I was expecting. p50 latency had gone from 40ms down to 4ms, CPU usage was sitting at a small fraction of where it used to be, and memory was barely registering at all. We had just finished moving the feature flags local evaluation endpoint from Django to Rust, and despite knowing that it was going to be better – that was the whole reason for this migration – I still wasn't ready for how the comparison ended up looking.
+
+## What local evaluation is


Suggested change

## What local evaluation is

## Local evaluation explained

jina-yoon · 2026-05-19T01:18:45Z

+
+The cache hit rate sits at 99.98%, which means almost every request is served from Redis with zero Postgres on the hot path. The dedicated Karpenter pool of memory-optimized instances is gone.
+
+I keep coming back to the rollout. Being able to shift 10% of traffic, sit with the metrics for a day, and roll back instantly if something looked off is what made this safe to ship fast. If we'd cut over in one step on day one, customer support tickets would have done the testing for me, and I would have been firefighting instead of writing this.


Suggested change

I keep coming back to the rollout. Being able to shift 10% of traffic, sit with the metrics for a day, and roll back instantly if something looked off is what made this safe to ship fast. If we'd cut over in one step on day one, customer support tickets would have done the testing for me, and I would have been firefighting instead of writing this.

---

The metrics are definitely eye-catching, but what I keep coming back to is the rollout. Being able to shift 10% of traffic, sit with the metrics for a day, and roll back instantly if something looked off is what made it safe to get there. It's also what let us catch three bugs while Django was still handling most of the load. If we'd cut over in one step on day one, it would have been customer support tickets doing the testing for me, and I would have been firefighting incidents left and right instead of writing this.

jina-yoon · 2026-05-19T01:44:18Z

+
+I keep coming back to the rollout. Being able to shift 10% of traffic, sit with the metrics for a day, and roll back instantly if something looked off is what made this safe to ship fast. If we'd cut over in one step on day one, customer support tickets would have done the testing for me, and I would have been firefighting instead of writing this.
+
+There's [Untangling Tokio and Rayon in production](/blog/untangling-rayon-and-tokio) if you want to read about an earlier optimization on the same Rust service, and [How we built the Rust feature flag service](/blog/feature-flags-service) for the original Django-to-Rust migration that this one builds on.


Suggested change

There's [Untangling Tokio and Rayon in production](/blog/untangling-rayon-and-tokio) if you want to read about an earlier optimization on the same Rust service, and [How we built the Rust feature flag service](/blog/feature-flags-service) for the original Django-to-Rust migration that this one builds on.

_If you want to read more about this topic, check out our blog on [Untangling Tokio and Rayon in production](/blog/untangling-rayon-and-tokio) which covers an earlier optimization on the same service, or [How we made feature flags even faster and more reliable](/blog/even-faster-more-reliable-flags) for the original migration that this one builds on._

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

patricio-posthog · 2026-05-19T12:19:06Z

+        "name": "Patricio Tarantino",
+        "role": "Product Engineer",
+        "link_type": "github",
+        "link_url": "https://github.com/patricio-posthog",


I noticed other people has a link to their LinkedIn, I don't mind, but if you want to keep it standard, this is my Linkedin https://www.linkedin.com/in/ptarantino/

patricio-posthog · 2026-05-19T12:20:19Z

Maybe @haacked wants to proof read it about the technical parts?

Also, @jina-yoon, when I go to the Blog in the deploy preview, it's broken (blue screen).

Add Patricio's blog post on Django-to-Rust local flag evaluation migr…

6651987

…ation Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

jina-yoon requested a review from patricio-posthog May 19, 2026 00:51

jina-yoon commented May 19, 2026

View reviewed changes

Comment thread contents/blog/feature-flags-definitions-rust.md Outdated

jina-yoon commented May 19, 2026

View reviewed changes

Comment thread contents/blog/local-flag-evals-rust.md Outdated

jina-yoon commented May 19, 2026

View reviewed changes

jina-yoon added 2 commits May 18, 2026 18:20

Apply suggestion from @jina-yoon

692d60d

Create local-flag-evals-rust.md file

fcb84f4

jina-yoon commented May 19, 2026

View reviewed changes

jina-yoon and others added 2 commits May 18, 2026 18:48

Apply suggestion from @jina-yoon

8b6c33f

Update title, add internal links to Patricio's local flag eval post

dc41851

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

patricio-posthog reviewed May 19, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add blog post: Moving local flag evaluation from Django to Rust#16882

Add blog post: Moving local flag evaluation from Django to Rust#16882
jina-yoon wants to merge 5 commits into
masterfrom
django-rust-ff-blog

jina-yoon commented May 19, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented May 19, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented May 19, 2026 •

edited

Loading

Uh oh!

jina-yoon May 19, 2026

Uh oh!

Uh oh!

Uh oh!

jina-yoon May 19, 2026

Uh oh!

jina-yoon May 19, 2026 •

edited

Loading

Uh oh!

jina-yoon May 19, 2026

Uh oh!

jina-yoon May 19, 2026

Uh oh!

jina-yoon May 19, 2026

Uh oh!

jina-yoon May 19, 2026

Uh oh!

patricio-posthog May 19, 2026

Uh oh!

patricio-posthog commented May 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

	metaDescription: "How we moved PostHog's feature flags local evaluation endpoint from Django to Rust, dropping CPU usage by 24x and memory by 56x."
	metaDescription: "How we moved PostHog's feature flags local evaluation endpoint from Django to Rust and dropped CPU usage by 24x and memory by 56x."


		---

		I reloaded Grafana three times before I trusted what I was looking at, because the gap between the old numbers and the new ones was bigger than I was expecting. p50 latency had gone from 40ms down to 4ms, CPU usage was sitting at a small fraction of where it used to be, and memory was barely registering at all. We had just finished moving the feature flags local evaluation endpoint from Django to Rust, and despite knowing that it was going to be better – that was the whole reason for this migration – I still wasn't ready for how the comparison ended up looking.

-I reloaded Grafana three times before I trusted what I was looking at, because the gap between the old numbers and the new ones was bigger than I was expecting. p50 latency had gone from 40ms down to 4ms, CPU usage was sitting at a small fraction of where it used to be, and memory was barely registering at all. We had just finished moving the feature flags local evaluation endpoint from Django to Rust, and despite knowing that it was going to be better – that was the whole reason for this migration – I still wasn't ready for how the comparison ended up looking.
+I reloaded Grafana three times before I could trust what I was looking at. I knew it was going to be better – that was the whole reason for this migration – but I still wasn't ready for how big of a difference it would make. p50 latency had gone from 40ms down to 4ms, CPU usage was sitting at a small fraction of where it used to be, and memory was barely registering at all.
+Here's how we moved our local flag evaluation from Django to Rust.

	Latency improved too. These numbers are measured at the Envoy layer (the time between Envoy receiving the request and getting a response back from the pod), so they don't include client-to-Envoy network time. A customer's SDK sees this plus its own network round-trip on top. Same Prometheus metric (`envoy_cluster_upstream_rq_time`) for both services, so it's apples to apples.
	Latency improved, too. These numbers are measured at the Envoy layer (the time between Envoy receiving the request and getting a response back from the pod), so they don't include client-to-Envoy network time. A customer's SDK sees this plus its own network round-trip on top. Since it's the same Prometheus metric (`envoy_cluster_upstream_rq_time`) for both services, it's apples to apples.


		I reloaded Grafana three times before I trusted what I was looking at, because the gap between the old numbers and the new ones was bigger than I was expecting. p50 latency had gone from 40ms down to 4ms, CPU usage was sitting at a small fraction of where it used to be, and memory was barely registering at all. We had just finished moving the feature flags local evaluation endpoint from Django to Rust, and despite knowing that it was going to be better – that was the whole reason for this migration – I still wasn't ready for how the comparison ended up looking.

		## What local evaluation is


		The cache hit rate sits at 99.98%, which means almost every request is served from Redis with zero Postgres on the hot path. The dedicated Karpenter pool of memory-optimized instances is gone.

		I keep coming back to the rollout. Being able to shift 10% of traffic, sit with the metrics for a day, and roll back instantly if something looked off is what made this safe to ship fast. If we'd cut over in one step on day one, customer support tickets would have done the testing for me, and I would have been firefighting instead of writing this.

	I keep coming back to the rollout. Being able to shift 10% of traffic, sit with the metrics for a day, and roll back instantly if something looked off is what made this safe to ship fast. If we'd cut over in one step on day one, customer support tickets would have done the testing for me, and I would have been firefighting instead of writing this.
	---
	The metrics are definitely eye-catching, but what I keep coming back to is the rollout. Being able to shift 10% of traffic, sit with the metrics for a day, and roll back instantly if something looked off is what made it safe to get there. It's also what let us catch three bugs while Django was still handling most of the load. If we'd cut over in one step on day one, it would have been customer support tickets doing the testing for me, and I would have been firefighting incidents left and right instead of writing this.


		I keep coming back to the rollout. Being able to shift 10% of traffic, sit with the metrics for a day, and roll back instantly if something looked off is what made this safe to ship fast. If we'd cut over in one step on day one, customer support tickets would have done the testing for me, and I would have been firefighting instead of writing this.

		There's [Untangling Tokio and Rayon in production](/blog/untangling-rayon-and-tokio) if you want to read about an earlier optimization on the same Rust service, and [How we built the Rust feature flag service](/blog/feature-flags-service) for the original Django-to-Rust migration that this one builds on.

	There's [Untangling Tokio and Rayon in production](/blog/untangling-rayon-and-tokio) if you want to read about an earlier optimization on the same Rust service, and [How we built the Rust feature flag service](/blog/feature-flags-service) for the original Django-to-Rust migration that this one builds on.
	_If you want to read more about this topic, check out our blog on [Untangling Tokio and Rayon in production](/blog/untangling-rayon-and-tokio) which covers an earlier optimization on the same service, or [How we made feature flags even faster and more reliable](/blog/even-faster-more-reliable-flags) for the original migration that this one builds on._

Conversation

jina-yoon commented May 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

To do before merging

Uh oh!

github-actions Bot commented May 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Deploy preview

Uh oh!

github-actions Bot commented May 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jina-yoon May 19, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

jina-yoon May 19, 2026

Choose a reason for hiding this comment

Uh oh!

jina-yoon May 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jina-yoon May 19, 2026

Choose a reason for hiding this comment

Uh oh!

jina-yoon May 19, 2026

Choose a reason for hiding this comment

Uh oh!

jina-yoon May 19, 2026

Choose a reason for hiding this comment

Uh oh!

jina-yoon May 19, 2026

Choose a reason for hiding this comment

Uh oh!

patricio-posthog May 19, 2026

Choose a reason for hiding this comment

Uh oh!

patricio-posthog commented May 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jina-yoon commented May 19, 2026 •

edited

Loading

github-actions Bot commented May 19, 2026 •

edited

Loading

github-actions Bot commented May 19, 2026 •

edited

Loading

jina-yoon May 19, 2026 •

edited

Loading