blog post: Surgery on a load-bearing assumption by dmarticus · Pull Request #16361 · PostHog/posthog.com

dmarticus · 2026-04-14T22:57:24Z

Changes

New engineering blog post: "Surgery on a load-bearing assumption"

An eng blog post about changing a foundational assumption in PostHog's feature flag evaluation engine (that a flag targets exactly one entity type) to support mixed person+group targeting. Structured as a reusable playbook:

The stent: adding a new field alongside the old one with a safe fallback (Option<Option<i32>> / effective_aggregation)
The staging: constraining first, then relaxing -- shipping the new data model with the old constraint enforced, then lifting it after the evaluator catches up
Post-op: hunting for the assumption in places that never referenced the field directly, like a reason-priority scoring function

Checklist

I've read the docs and/or content style guides.
Words are spelled using American English
Use relative URLs for internal links
I've checked the pages added or changed in the Vercel preview build
If I moved a page, I added a redirect in vercel.json

github-actions · 2026-04-14T22:57:35Z

Deploy preview

Status	Details	Updated (UTC)
🟢 Ready	View preview	Apr 21, 2026 10:41PM

github-actions · 2026-04-14T22:58:17Z

Vale prose linter → found 0 errors, 4 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/mixed-targeting-feature-flags.md — 0 errors, 4 warnings, 0 suggestions

Line	Severity	Message	Rule
22:43	warning	Capitalize 'Feature Flags' for PostHog's product. Use 'feature flags' for the general industry concept.	`PostHogBase.ProductNames`
50:28	warning	Capitalize 'Feature Flags' for PostHog's product. Use 'feature flags' for the general industry concept.	`PostHogBase.ProductNames`
72:155	warning	'serde's' is a possible misspelling.	`PostHogBase.Spelling`
72:213	warning	'deserializer' is a possible misspelling.	`PostHogBase.Spelling`

dmarticus · 2026-04-14T22:58:29Z

@@ -0,0 +1,121 @@
+---


qq: should I also name this file the same name as the post, e.g.

mv contents/blog/mixed-targeting-feature-flags.md contents/blog/surgery-on-a-load-bearing-assumption.d

?

we can go with the title but maybe just: load-bearing-assumption-surgery?

We're not getting any SEO points for any of these so can be just short and close to the title.

ivanagas

Higher level thoughts, can do a more detailed pass after you update

ivanagas · 2026-04-15T20:19:09Z

@@ -0,0 +1,121 @@
+---
+title: "Surgery on a load-bearing assumption"


I know people around here hate stuff like this, but:

Suggested change

title: "Surgery on a load-bearing assumption"

title: "How to do surgery on a load-bearing assumption"

As a reader, makes me a bit more curious, I might have load-bearing assumptions I want to do surgery on too.

I'm okay with "How to" but i think having both "surgery" and "load-bearing" is too many metaphors in one title. more variations -

Suggested change

title: "Surgery on a load-bearing assumption"

title: "How to perform surgery on a live assumption"

Suggested change

title: "Surgery on a load-bearing assumption"

title: "How I performed surgery on a load-bearing assumption"

title is hard lol i'm not sure about any of these yet either

I don't mind "How I performed surgery on a load-bearing assumption"; I'll make that change.

ivanagas · 2026-04-15T20:23:03Z

+  metaDescription: "How we changed a foundational assumption in PostHog's feature flag engine -- and a playbook for doing it in your own systems."
+---
+
+I recently shipped [mixed targeting](/docs/feature-flags/user-and-group-targeting) for [feature flags](/feature-flags) -- letting a single flag target both users and groups at the same time. 15 PRs across four repos. The part I keep thinking about is how hard it was to find all the places the old assumption lived.


In this intro, before the json code, we want to:

Make it clear what we're talking about when we say "assumption"

Introduce the surgery analogy

If you are coming into this blog without context, this is unclear. Even after reading the full first section, I get that aggregation_group_type_index some data that groups inherit but I don't fully connect why that's an "assumption" or why you'd do surgery to get rid of it.

If we get this clear, there's not that much more that needs to be done.

i tried to do a rewrite and restructure of the intro to add all this context - feel free to add further edits!

ivanagas · 2026-04-15T20:24:18Z

Overall, I think there's some issue with whether we are using past tense or present tense, 1st person or 3rd person. I get edit this more closely on next pass, just something to be aware of.

+1! i just finished a pass and tried to fix them

ivanagas · 2026-04-15T20:25:23Z

@@ -0,0 +1,121 @@
+---


we can go with the title but maybe just: load-bearing-assumption-surgery?

We're not getting any SEO points for any of these so can be just short and close to the title.

ivanagas · 2026-04-15T20:26:10Z

+
+A three-variant enum would be more readable, but the JSON wire format is already `null` or a number, and `Option<Option<i32>>` maps to that directly with serde's defaults. A custom enum would need a hand-written deserializer to produce the same JSON shape, which is more code for the same result. Ugly, but the path of least resistance when you're keeping an existing API contract intact.
+
+The resolution logic is simple:


Can you explain what this logic does in like one sentence for us who don't know rust?

+1 i didn't touch most of the Rust explanations because I couldn't understand them

ivanagas · 2026-04-15T20:32:30Z

+
+You can grep for a field name. You can't grep for an assumption.
+
+## The stent


I know you like to keep your titles more stylish, but I literally didn't know what a stent was until I read the rest and figured it out, could it be a bit more informative?

Suggested change

## The stent

## Insert the stent

ivanagas · 2026-04-15T20:33:56Z

+
+The stent is the most important part of the whole project. It's what makes the rest of the surgery safe.
+
+## The staging


I don't really understand what's staged here? What would be the verb we would use with staging?

ivanagas · 2026-04-15T20:35:19Z

+
+The whole point is that the last incision is tiny. All the risk was managed in the earlier steps.
+
+## Post-op


I don't think this is the right way to say this but can we say something like this?

Suggested change

## Post-op

## Fixing new issues in post-op

ivanagas · 2026-04-15T20:37:07Z

+
+I found a few more places like this. The blast radius endpoint returned either user counts or group counts, never both. The frontend rendered one unit label per flag. The SDK request payloads assumed one hash key type. None of them referenced the field. All of them depended on the assumption.
+
+In my experience, assumptions tend to hide in three places: **ranking and priority logic** (like the scoring function -- anything that defines an ordering over outcomes), **API response shapes** (like the blast radius endpoint -- anything that assumes a fixed cardinality in its return type), and **UI rendering code** (like the unit labels -- anything that maps data to a display format). If you're doing this kind of surgery, those are the places I'd audit first.


For most readers, this is probably the most useful piece of information, but it's sort of buried here at the end. I'd love something like this at the beginning, even the beginning of the section.

jina-yoon · 2026-04-21T22:25:43Z

+author:
+  - dylan-martin
+featuredImage: >-
+  https://res.cloudinary.com/dmukukwp6/image/upload/580944574_1ff3a5f8_0cc0_4a25_8ba5_7e532793c1fb_0b30cb5f50.jpg


updated to the doctor hogs image, so cute

I'd die for Dr. Hog

jina-yoon · 2026-04-21T22:27:53Z

+
+This stent was the most important part of the whole project. It's what makes the rest of the surgery safe and possible.
+
+## Part 2: Staging


+1 to ian's comment earlier - something parallel to "Placing the stent" would be ideal here for staging

ah, I was trying to use "staging" to continue the surgery metaphor, but I'm fine to change it up here.

jina-yoon · 2026-04-21T22:28:55Z

+
+- The blast radius endpoint returned either user counts or group counts, never both. 
+- The frontend rendered one unit label per flag.
+- The SDK request payloads assumed one hash key type. 


these three bullets were not clear to me - could you flesh them out?

that way also they're not identical in format to the final bulleted list in the Lessons learned section.

jina-yoon · 2026-04-21T22:30:11Z

+2. **API response shapes** like the blast radius endpoint – anything that assumes a fixed cardinality in its return type
+3. **UI rendering code** like the unit labels – anything that maps data to a display format. 
+
+For anyone else out there doing this kind of surgery, those are the places I'd audit first.


+1 to ian's earlier comment - a bit more detail or reflection would be good to wrap up this blog post.

based on how we flesh out stuff in the previous section, i might agree that it would be best to bump these "lessons learned" to the beginning for anyone who wants to get the advice without reading the full story

dmarticus · 2026-04-22T16:32:33Z

@jina-yoon @ivanagas y'all are the best – thanks for the feedback! I'll wrap this up today.

abusing the ever-living fuck out of a metaphor

10223f1

dmarticus commented Apr 14, 2026

View reviewed changes

ivanagas reviewed Apr 15, 2026

View reviewed changes

jina-yoon mentioned this pull request Apr 16, 2026

Art request - Hero image request reusing doctor Max from hoggies #16413

Closed

jina-yoon and others added 2 commits April 21, 2026 15:22

significant restructure/rewrite

3858c08

Fix typos

09a09da

jina-yoon reviewed Apr 21, 2026

View reviewed changes

Merge branch 'master' into dmarticus/blog-idea-review

e693cd1

		@@ -0,0 +1,121 @@
		---
		title: "Surgery on a load-bearing assumption"

	title: "Surgery on a load-bearing assumption"
	title: "How to do surgery on a load-bearing assumption"

	title: "Surgery on a load-bearing assumption"
	title: "How to perform surgery on a live assumption"

	title: "Surgery on a load-bearing assumption"
	title: "How I performed surgery on a load-bearing assumption"


		A three-variant enum would be more readable, but the JSON wire format is already `null` or a number, and `Option<Option<i32>>` maps to that directly with serde's defaults. A custom enum would need a hand-written deserializer to produce the same JSON shape, which is more code for the same result. Ugly, but the path of least resistance when you're keeping an existing API contract intact.

		The resolution logic is simple:


		You can grep for a field name. You can't grep for an assumption.

		## The stent


		The stent is the most important part of the whole project. It's what makes the rest of the surgery safe.

		## The staging


		The whole point is that the last incision is tiny. All the risk was managed in the earlier steps.

		## Post-op


		I found a few more places like this. The blast radius endpoint returned either user counts or group counts, never both. The frontend rendered one unit label per flag. The SDK request payloads assumed one hash key type. None of them referenced the field. All of them depended on the assumption.

		In my experience, assumptions tend to hide in three places: ranking and priority logic (like the scoring function -- anything that defines an ordering over outcomes), API response shapes (like the blast radius endpoint -- anything that assumes a fixed cardinality in its return type), and UI rendering code (like the unit labels -- anything that maps data to a display format). If you're doing this kind of surgery, those are the places I'd audit first.


		This stent was the most important part of the whole project. It's what makes the rest of the surgery safe and possible.

		## Part 2: Staging

Conversation

dmarticus commented Apr 14, 2026

Changes

Checklist

Uh oh!

github-actions Bot commented Apr 14, 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 Apr 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

dmarticus Apr 14, 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!

Choose a reason for hiding this comment

Uh oh!

ivanagas left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jina-yoon Apr 17, 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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jina-yoon Apr 21, 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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

jina-yoon Apr 21, 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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dmarticus commented Apr 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

github-actions Bot commented Apr 14, 2026 •

edited

Loading

github-actions Bot commented Apr 14, 2026 •

edited

Loading

dmarticus Apr 14, 2026 •

edited

Loading

jina-yoon Apr 17, 2026 •

edited

Loading

jina-yoon Apr 21, 2026 •

edited

Loading

jina-yoon Apr 21, 2026 •

edited

Loading