Re: LLM Usage Policy

Sat, 26 Sep 2026 10:51:22 -0700

I’m fine with rolling my proposal into Blake’s, with the understanding that we still tag model assists in our commit messages.

On Sep 26, 2026, at 11:58 AM, Dinesh Joshi <[email protected]> wrote:


My +1 was to keeping it simple and just go with Blake's simple straightforward guidelines.

On Sat, Sep 26, 2026 at 9:47 AM <[email protected]> wrote:
We have multiple active proposals on this thread but it’s unclear which proposals "+1’s” should be attributed to.

From a threaded reading based on which messages reply to which:
  • Alex’s reply with a +1 is to Benedict’s email and original proposal.
  • Jon’s reply is to Blake’s email but says “to simple straightforward guidelines” (unclear if referring to Blake’s proposal or simple guidelines in general).
  • Dinesh’s reply is to Jon’s, which carries ambiguous attribution to Blake’s proposal or simple guidelines.

To keep our discussion tidy and positions clear, can I propose clearly attaching a +1 to a specific proposal or position in the thread so there’s no ambiguity as to what each of us support?

– Scott

On Sep 26, 2026, at 9:32 AM, Dinesh Joshi <[email protected]> wrote:

+1

On Sat, Sep 26, 2026 at 9:21 AM Jon Haddad <[email protected]> wrote:
+1 to simple straightforward guidelines 



On Thu, Sep 24, 2026 at 1:01 PM Blake Eggleston <[email protected]> wrote:
I propose the following guidelines:

  1. Contributors must have a comprehensive understanding of the patches they contribute
  2. Reviewers must have a comprehensive understanding of the patches they approve
  3. Reviewers must have experience with the affected systems commensurate with the risk and complexity of the patch.

This makes the expectations that have been mostly implicit explicit, doesn’t gate keep on the contributor side, doesn’t tell people how to do their work, and gives the subject matter expert, the reviewer, the latitude to decide what standards make sense for the patch they’re reviewing. 


On Thu, Sep 24, 2026, at 12:01 PM, Josh McKenzie wrote:
Apache Cassandra was fundamentally undeployable for four years between Nov 2015 - 2019.
CASSANDRA-8099 was a maximal manifestation of a specific approach to engineering and calendar constraints we've seen time and again on the project; I don't want us to conflate things here. That was a herculean monolithic body of work performed in inhuman conditions (in vim!) that was ultimately so invasive, all the unit tests in the code-base were commented out and Jake and I spent a grueling 1.5-2 months hand-rewriting basically all the unit tests in that code-base to get things to even build and run, much less pass.

Massive blast radius changes that are un-sustainably complex, under-tested, where we don't property, fuzz, check coverage, check complexity, or A/B compare against a known good system (i.e. pre/post) correctness testing are going to destabilize the database at any time under any regime of tooling. We certainly could speed-run our way back into destabilization with LLM's as they are a force multiplier for both the good and the bad of one's engineering practices, but that's a solvable problem by better defining what our bars of quality are (Definition of Done anyone?) and holding ourselves accountable to delivering at that bar. 

On Thu, Sep 24, 2026, at 2:40 PM, David Capwell via dev wrote:


if you really want to pursue it I would ask that we do it offline to avoid polluting an already busy conversation

People are directly responding saying that they feel discrimination currently and that the policy tries to codify that discrimination, so I feel its 100% on topic. This thread has presented 0 evidence that LLM usage has lowered the quality of contributions merged and has so far been vibes and feeling; I have yet to see any evidence to justify such discrimination so I will keep pushing back until such evidence is presented so we can have a informed debate.

namely how we handle shallow and localised bug fixes. I would be happy adding a clear entry to the “Permitted” section for this. No doubt there are many refinements needed to the Restricted text as well, that might also capture some of your concerns.

I will not sign off on cherry picking areas where "safe" to use a tool; so no it does not capture my concerns.


I don't know if everyone remembers, but ten years ago Cassandra was full of serious correctness and stability issues. Despite developing it, I would not have run it myself or recommend that anyone use it. We have dug ourselves out of that hole, but it took years of discipline and effort, and we're still (deservedly) recovering our reputation. 

Let's use this new technology to improve the quality of our contributions, not squander our hard-earned gains in the name of speed. It will be hard to recover our reputation a second time.

I do recall the 3.x line and put in a significant amount of effort to harden it. There were behaviors I noticed after joining Cassandra that I feel directly contributed to 3.0 and the decade of catch up; behaviors that still linger in parts of the community today.

As I look on trunk and look at committed code and trace back to PRs and JIRA I see the following:

  • large patches approved without comments
  • 0 evidence that tests were run

I then look at our CI and see tests failing for months. As you start to triage you start to see some of them show real issues; yet they linger for months not being addressed... when CI is unstable it takes a lot of effort to triage "did my patch break the test", and I have seen time and time again people do not put in that effort, and shrug off as "its just a flakey test"; then our CI failure rate grows.

Non of this has anything to do with LLMs but LLMs running in this environment is far more dangerous as there are not checks in place to "hold the bar". I am all for raising the bar universally; expecting both humans and LLMs to match that bar.

I’d support something that boils down to roughly this:

1.) 2 committers must understand an LLM-assisted change before it commits. (Perhaps separately we can explore the question of why we haven’t added any new committers to the core project for about a year. I’m also still not entirely sure if it’s acceptable within our guidelines for a committer to +1 a patch after delegating review.)

2.) Patch authors must demonstrate enough understanding to discuss their own patch, whether or not parts of it are generated by an LLM.

3.) The “Assisted-by” tag should be used to indicate any non-trivial LLM usage in the generation of a patch, just like we have used Co-authored-by historically.

4.) Comments and other things that aren't the actual code (but could sow confusion) should be held to the same standard we'd expect from a human writer. If we don’t yet agree on that standard, we can formalize enough of it to guide both humans and LLMs.

I can get behind this proposal but i would tweak it as 1/2 i don't think really need to special case LLM usage

  1. 2 committers must understand the change before it commits.
  2. Patch authors must demonstrate enough understanding to discuss their own patch

Nothing about those 2 need to be scoped to LLM usage and honestly matches most PMCs I have talked to understanding of our bar (as Benedict pointed out, the actual wording could be interpreted to allow rubber stamping from committer)

As for 3 I am cool with this. ASF recommends the same (it says Generated-by but that discount's the human's effort) as its useful for audits and tooling. Having Assisted-by tag should not imply anything about the committed patch as it should have gone through the same bar we all expect; its just for tools auditing.


On Sep 24, 2026, at 10:37 AM, Aleksey Yeshchenko via dev <[email protected]> wrote:

Meant "doing away with", sorry. Non-native speaker with a headache here. Thanks Caleb for spotting.

On 24 Sep 2026, at 17:35, Aleksey Yeshchenko via dev <[email protected]> wrote:

P.S. I assume it's obvious from the text above that I don't believe that getting away with human code review is a viable option.


On 24 Sep 2026, at 17:37, Štefan Miklošovič <[email protected]> wrote:

Good call on checkerframework, we even have a patch for it. Work of
Jacek Lewandowski. We might just drive it to completion. Using AI for
finishing it would be quite ironic.


On Thu, Sep 24, 2026 at 6:19 PM Jon Haddad <[email protected]> wrote:

There are some really good points being brought up about stability of the codebase, maintainability, quality of reviews, correctness bugs, and I agree with all of them.  I think it would be helpful to take a step back and consider how those bugs got there in the first place, how they were fixed, and what we could do to further advance the codebase so they don't creep back.  LLMs can be used either with very tight guardrails, or in what's effectively YOLO mode, and there's a big difference in the quality of the results you get.

One thing to keep in mind, a lot of the initial code in C* was added without comprehensive testing.  I hope we can all agree that it's a lot easier to break code that doesn't have high quality tests.  During the code freeze, a lot of people people spent several years relentlessly finding and fixing bugs. This was probably a pretty frustrating time for anyone who was focused on fixing other people's bugs when they wanted to build features.  I think we should recognize the effort here and appreciate the foundation that the project stands on now. I can understand how anyone involved with this effort would be apprehensive about seeing years of their life swept away by an agent that was driven by goal seeking to remove all the tests that it broke instead of fixing them.

When I picked up the work to improve cursor compaction, the first thing I asked myself was how can I make sure I don't break this?  How do I even know it works properly?  There were some tricky parts to the code, and I really didn't want to come in and immediately break stuff.  That's why I started with an entire patch dedicated to adding test infra to it.  90% of the patch was tests, and in my other cursor patches, it remains *at least* 80% of my patches.  It was a *lot* faster to add almost 10K lines of tests that handled a byte for byte differential testing paired with harry to find over 30 bugs that caused cursor to corrupt results.  Range tombstones alone were at least a dozen bugs, but I also found issues with static columns, reverse ordering, etc.  Randomizing schemas and data in burn tests to generate different shapes of data, to ensure they all result in the same output at the end. JMH tests to ensure there weren't performance regressions, hours of profiling. These were all a *lot* easier to do with the LLM helping me out.  In the process I've found bugs that have been lingering in the codebase for years.

That's a long story, but hopefully we all agree that having comprehensive tests is a great way to ensure that both humans and LLMs don't break things that are working.

The lesson: we need to keep improving our testing.  Everything that we touch, should be left in a better state than how we found it with regard to test coverage.

Test coverage isn't everything though, there's always little subtle bugs that don't get found in testing, that can slip in despite our best efforts.  It's debatable if humans will be as good as agents for coding in the long term, for spotting small defects.  I sincerely doubt it.  For the time being though, we still have people involved. It's probably a good time to start using more static analysis tools to identify problematic code and to add this to CI.  Dmitry had a suggestion recently for checkerframework to detect leaking contexts, a problem he spotted when reviewing my branch.  It would be great to have that integrated into our CI and dev workflow so we can simply avoid an entire class of bugs.

There's also PMD, which is excellent for finding code that can be hard to understand.  I *highly* suggest you all run PMD to analyze for cognitive complexity and high npath scores.  This was made popular by the folks at Sonar and I've found it to be an excellent feedback mechanism for structuring code.  The default max they set is 15, which is the point where it starts to become difficult to verify something works without making a massive investment.  We've got areas in the codebase that are in the hundreds, and some parts even higher.  These have been contributed by humans, and are all high risk points for both humans and agents to start messing around with.  They're also in some fairly critical areas that are very likely to break, so I understand why people would not want an agent anywhere near it.

Unfortunately, it's not an easy problem to address.  There's so many places where the code is structured in a way that has so many branches, so many conditions, that it's effectively impossible for a human to understand, creating a fear of messing around in it.  There's plenty of areas that deserve extreme scrutiny, and we should be careful of what we add, whether it's human or agent.

The codebase today requires a high degree of internal knowledge to navigate.  There's land mines everywhere. We should be looking to make conscious improvements by moving the code forward, so it's easier to make changes to small, well tested components with minimal side effects.  Not making it harder for people to use the tools that aid in that process.

Here's what we could do to achieve the underlying goal of not breaking the DB:

Add cognitive complexlity and npath via PMD as a feedback mechanism.

Code that's hard to understand is hard to review.  It's also hard to test. Let's break down the complex code so more people can contribute, safely.

Add checkerframework to our tooling,

Properly annotate the codebase for it and reduce the surface area that things can break.  Less brittle codebase = "" can move faster.

Use jacoco to find areas of the codebase with poor testing.

Let's improve the test coverage there, LLMs are great for this.  We have a ton of static tests, these can become more dynamic, parameterized, and leverage harry.

Refactor parts of the codebase that have high cognitive complexlity and NPath scores.

This should be lowered over time to meet some high watermark, say 25 maximum, although I'd prefer 15 which is where the Sonar folks settled.

Move forward moving the codebase to a more modular structure

We've talked about Gradle on and off - but it can really be a huge help with incremental, modular builds. This is pretty easy to do with an agent and we could have it done in a couple days.

Enforce boundaries with ArchUnit

If we want to enforce certain code boundaries, this is the way to do it. Should not be part of manual review.

Add LLM review for all incoming PRs before a human

The goal here is to automate the initial part of the review process that reviewers should spot, and raise the bar for the initial contribution.  When the code gets reviewed by a human, it should already have passed a large variety of initial checks.  This should shorten the review cycle and result in higher quality patches.  I've had Claude reviewing all my PRs in my personal projects for a while now and it consistently gives great feedback that I almost always incorporate.

In my ideal world, we'd also auto-format all code

Consistent formatting throughout the codebase would be amazing, but that's just one man's dream.

Hopefully there's at least a couple things in this list we could move forward with in the short term, as it'll help improve the code quality regardless of how it's created.

Jon






On Thu, Sep 24, 2026 at 7:38 AM C. Scott Andreas <[email protected]> wrote:

From Benedict:

“I don't know if everyone remembers, but ten years ago Cassandra was full of serious correctness and stability issues. Despite developing it, I would not have run it myself or recommend that anyone use it. We have dug ourselves out of that hole, but it took years of discipline and effort, and we're still (deservedly) recovering our reputation.”

Expanding on this point for those who may not have been active in the project at this time —

Apache Cassandra was fundamentally undeployable for four years between Nov 2015 - 2019. The database literally lost data if you ran a read-only SELECT query ordered descending (C-14513, C-14515). If you haven’t read these tickets before, please take a moment to do so.

It took years of careful work via property-based testing, fuzzing, and deterministic simulation to restore Cassandra’s status as a usable system of record. Once 14513 and 14515 were identified, nearly 30 additional critical data loss and incorrect response bugs were identified.

It is essential for the project’s future that we don’t regress to this state chasing AI-generated features motivated by fear. The fact that examples cited in this thread which boast shiny features but have critical shortcomings unknown to their author supports this argument.

The most common path for large corpuses of AI-generated software is elation and reveling in a feature matrix, followed by abandonment.

I endorse this point:

“Let's use this new technology to improve the quality of our contributions, not squander our hard-earned gains in the name of speed. It will be hard to recover our reputation a second time.”

Patrick, I don’t want your note regarding a TCM issue to go unaddressed. Please file a Jira ticket and the patch if you like. I can’t comment on the patch as I haven’t seen it, but together we will solve the problem.

– Scott

On Sep 24, 2026, at 4:01 AM, Benedict Elliott Smith <[email protected]> wrote:

Hi Patrick,

As I mentioned in my reply to David, I would be happy to create a carve out for shallow and localised bug fixes in the "Permitted" section. Would this alleviate some of your concerns regarding your ability to contribute to the project?

I appreciate your pointing out Ferrosa's Accord implementation however, as it is a *great* example of the problems we're leaping into. I took a look, and within about 30s found that the protocol is fundamentally incorrect, having failed to address CASSANDRA-18365. This is despite claiming to be tested with Jepsen that should in principle find this fault. I followed up by using Claude to interrogate the implementation further, and immediately found other serious correctness issues.

I use LLMs daily now to help facilitate Accord development, and while they are powerful they are NOT able to author the code themselves, even when building upon a strong human-authored foundation.

I don't know if everyone remembers, but ten years ago Cassandra was full of serious correctness and stability issues. Despite developing it, I would not have run it myself or recommend that anyone use it. We have dug ourselves out of that hole, but it took years of discipline and effort, and we're still (deservedly) recovering our reputation.

Let's use this new technology to improve the quality of our contributions, not squander our hard-earned gains in the name of speed. It will be hard to recover our reputation a second time.


On 2026/09/23 19:16:17 Patrick McFadin wrote:
I was waiting for this moment to hit our project and I'm glad we're here. I
am deeply concerned for our project and its future, as we have increasingly
made it difficult to contribute. I had hoped that this new era of
software tools powered by AI would expand the project's reach and bring
more diverse thoughts and ideas. This policy proposal is the exact opposite
of what we need. We have been sitting on a Cassandra 6 release alpha for
months. We need to accelerate and embrace new ways of being or be left
behind. As I read that policy, my first and gut level reactions:
- It comes across as elitist and class protectionism. Committer should not
be special but this proposal makes that designation even more sacred.
- It signals that our project is so fragile that only a few people "Really
understand it" That's some SQLite vibes right there.
- Trying to fix a problem that doesn't exist
Sadly, i think this policy change would also exclude a lot of comitters.
We aren't alone in this moment. The Linux project just went through
this. You can find the thread with a simple Google, but similar hard
feelings were being expressed "AI is going to ruin our project!", "The
unwashed masses are going to contribute terrible code!", "We have to
protect our precious status as Linux maintainers!"  Linus being Linus, was
deeply invloved and they adopted a super simple statement that covers all
bases. Human or Human using AI. “You are expected to understand and to be
able to defend everything you submit.”  Love that.
In the larger picture, I'll restate. I'm worried for our project. In late
2025(Opus 4.5 IYKYK), early 2026, AI coding LLMs turned a real corner and
in the hands of somebody that knows how to build software, this tool is
like jet fuel. Here's some examples of new projects being hyper fueled by
AI coding tools.
Apache Iggy - Complete rust replacement of kafka. Crazy fast velocity
Turso - Rust re-write of SQLite
Bun - Rust re-write of itself from Zig.
Think this couldn't happen to us? Already has:
https://github.com/ferrosadb/ferrosa. Ben is using it to power his own
startup, but it was him alone using a ton of local AI coding agents. He
even implemented Accord. Yeah...
The cracks are already starting to show. There is a black market economy of
Cassandra patches happening now. Not going to name names or call people
out,  but there are fixes and optimizations living in branches outside of
the Cassandra project. Why? I'll use myself as an example. I fixed a nasty
bug I ran into with TCM a few weeks ago. Wrote the tests. It passes CI and
lives in my personal branch. I'm sitting here really wondering if I want to
go through the ritual humiliation of being roasted for using AI to fix it.
Me. I am worried about contrinuting code the Cassandra. What the hell does
that say?
I have my CQLite project that I've been doing a release around once a
month. I would love to donate that to the Cassandra project but I wouldn't
if it essentially killed any progress.
My larger counter proposal would be to:
- Adopt the “You are expected to understand and to be able to defend
everything you submit.” approach the Linux project has adopted.
- Loosen up the contributor process and our worry on trunk. Let 1000
flowers bloom and bring it in.
- And finally, to give some people more peace of mind and open more doors,
adopt what other projects have done and provide more pluggability. Let new
ideas have an easy place to connect.
We are at a fork in the road. What are we going to do? And then I have to
ask myself, what am I going to do as a contributor?
Patrick
On Wed, Sep 23, 2026 at 6:16 AM Blake Eggleston <[email protected]>
wrote:
I’m not necessarily opposed to having a policy, but so far we have some
specific proposals addressing a problem statement that’s very nebulous.
What is the community failing to do on its own that we’re trying to correct
with policy? What outcomes are we trying to create or prevent? Having some
examples and specific problems to discuss would help focus the conversation.
On Wed, Sep 23, 2026, at 4:34 AM, Shailaja Koppu via dev wrote:
Benedict,
Thanks for clarifying. My concern still remains. This criteria would be
difficult to define and apply consistently. What counts as “similar” scope
or area, “mostly correct,” or sufficiently independent work? More
importantly, how do we prevent such vague criteria from creating an
informal hierarchy where some contributors work is routinely accepted while
others is routinely rejected?
If the intent is to limit AI-assisted code changes to Cassandra
contributors, or to contributors who have previously worked in that
component without AI, that would at least be clear and enforceable.
On Sep 23, 2026, at 12:01 PM, Benedict Elliott Smith <
Core code changes
Chris: Do you object to the first or second line you quote? Because the
first line is effectively motivation for the second line, and can be
removed (or more clearly combined). If it’s the second line, then I do not
think this is an unreasonable expectation, and we can get into a proper
debate about it.
Shailaja, since you only snipped the first sentence, your concerns might
also be mostly answered by this clarification? “Minimal third-party
guidance” implies you have some concerns about the second line, but all of
our policies have some ambiguity because legalese is even worse. I don’t
think the ambiguity here would be challenging to navigate though we can
certainly refine it. This specific snippet is meant to convey an
expectation that a contributor has autonomously produced patches of similar
scope that were mostly correct, so that they have demonstrated the level of
understanding necessary to guide another party to a successful patch (i.e.
an LLM in this case).
On 2026/09/23 10:54:16 Benedict Elliott Smith wrote:
Thanks everyone for your input so far. I’ll respond in brief to the
main themes, in (mostly) separate emails so they can each have their own
debate chain.
Should we have a policy (Blake/Josh*/Jon/Dinesh)
I think we would all agree that LLMs represent the biggest change to
this community (and software more generally) since its inception, and we
all now have enough experience with the technology to have formed opinions
about how it is best managed. We also evidently have not all arrived at the
same conclusions. In this situation, it would be an abdication of our
responsibilities as a management committee to not agree *some* policy.
I intend to conduct straw polls as the discussion evolves, so if you
prefer an alternative policy - or modifications to this policy - I would
encourage you to make those alternative proposals.
*Veto/Consensus (Josh)
It was fair to call out my poor use of language on this topic, so let
me rephrase a little. The community is built on consensus, and work should
not be merged when there are outstanding concerns to address. The explicit
-1 should only be used rarely, because the prior expectation should prevent
it ever being needed. I (and others) have outstanding concerns on LLM
generated work that can only be addressed through this process right here,
so to merge such work while maintaining the community’s consensus we must
agree some policy.
On 2026/09/23 09:58:27 Shailaja Koppu via dev wrote:
I am strongly -1 on this
- Core code changes made by LLM may only be proposed by contributors
with demonstrated expertise
That creates a new, subjective privileged class of contributors and
turns a tool choice into an eligibility test. Who decides whether expertise
has been “demonstrated,” what counts as “minimal third-party guidance,” and
how could those judgments be applied consistently or fairly?
Apache already has a better model, anyone may contribute, trust and
additional repository privileges are earned transparently over time. The
ASF describes its communities as flat, and says that newcomer ideas have as
much input as those from original creators. We should not add a separate,
informal hierarchy in which certain people may use common development tools
while others may not.
On Sep 23, 2026, at 6:33 AM, Chris Lohfink <[email protected]>
wrote:
- Core code changes made by LLM may only be proposed by contributors
with demonstrated expertise
- Must have produced similar patches in size, scope and area
unassisted and with minimal third-party guidance
I really don't like this one or its wording. Definitely too "the
peasants are getting uppity lets build a wall". Lets not let a subjective
thing like demonstrated expertise (who decides that?) be if it's ok or not.
Hold the same standards for code quality and process for it all. I don't
want this to be: only people on the storage team in Apple can use AI.
Chris
On Wed, Sep 23, 2026 at 12:16 AM <[email protected] <mailto:
I agree with Stefan and think this is both a reasonable and
thoughtful proposal.
Here are some things I like about it:
– It outlines areas where LLM usage is unambiguously useful to the
project’s developers and users.
– It defines a spectrum of recommendations and cautions.
– The only prohibited areas are extremely narrow and say nothing
about code at all.
Some in this thread are responding as if this proposal seeks to
prohibit or sharply limit use of LLMs. In fact, it’s one of the most open
and welcoming I’ve seen for an OSS project of our size where many are
adopting policies that simply ban them entirely. I’ve re-appended the
proposal below my message as it seems to have been lost in threaded
replies, and would encourage folks to give it a second read.
Some brief thoughts based on my own use of LLMs:
– I find them fantastically useful for reviewing and identifying
problems that have slipped through review - primarily via Alex Petrov’s
/deep-review skill, which I have running in a VM in a loop executing over
every new commit in the project as of a few days ago. I will be posting a
few hand-authored Jira tickets based on findings that appear legitimate to
me. For now, the loop is posting them as issue drafts for my own review on
my personal fork which you can find here:
– They’re great for enabling use of model checkers and formal
methods where such work would have previously been prohibitively expensive,
such as Blake’s work on a TLA+ proof of aspects of Mutation Tracking and
Benedict/Fedor’s work on a machine-checkable proof of the Accord protocol
in Lean.
– They are stunning for allowing me to experiment with ideas that
would have otherwise been a summer internship’s scope of work. Some
examples include an io_uring prototype, exploring the impact of
page-aligned compressed chunk sizes, an API shim bridging the 3.x and 4.x
Java Drivers, and potential enhancements to Zstandard.
– And they shine when given grunt-work that is critical to the
project but a miserable labor for humans, such as triaging, reproducing,
and root-causing flaky tests, which David Capwell now has running in a loop
to help us improve CI stability in the project.
I never thought I’d be so positive on what’s possible via language
models a year ago. At the same time, I also agree that they present
challenges and risks that can be managed through thoughtful discussion and
policy. Some of the concerns that I think are important to guard against
include:
– Asymmetry of effort between author and reviewers: As
token-generating machines, LLMs can generate diffs of extraordinary size
very rapidly. /deep-review is great for chewing through diffs and
identifying defects. But it should be used by the contributor themselves to
identify issues – not to replace the role of the reviewer with more
electricity. The role of the reviewers extends beyond identifying and
highlighting defects. It encompasses architecture, harmony with the
existing codebase, thinking ahead to future evolution of the project, and
replicates context on the project as new code is committed. These functions
cannot be automated away.
– Hesitancy of authors to engage manually with code they have
generated: This is not specific to Cassandra, but it is a behavior that I
have seen in several “highly-electric” projects. There’s a bimodal tendency
toward code that is entirely generated or entirely human-authored - but it
is rare for someone to prepare an AI-authored patch to take an offramp and
spend a significant amount of time refining the work by hand in an IDE.
This hesitancy toward human participation in authorship of LLM-generated
code is very concerning to me.
– Harmony with the existing codebase: Due to the tunnel-vision of
context windows, LLMs are generally unaware of conventions and norms
present in codebases and very frequently reinvent concepts in a generation
turn to suit a goal without view of the project’s overall architecture.
This results in a profusion of messy and duplicated concepts that gradually
sprawl about a codebase.
Again, none of these are grounds for prohibition of usage of
language models in developing the project. They’re just problems we need to
bear in mind and guard against – and I think the proposal is designed to do
just that.
I’m thrilled by the potential of LLMs to improve Apache Cassandra
and we already see it happening through a vast number of issues that are
being reported and fixed. But there’s also danger in taking ATVs down a
hiking trail full of people.
Regarding the prohibition on prose, I’ll simply say: I recently
found myself in a scenario where I found a Claude-authored document so
inscrutable that I piped it back into a model, directed it to rewrite it in
ASD-STE100, read it myself, and responded based on the summarization. As a
humanities grad, this is probably the worst language crime I have
committed. But it was in response to language that was itself so
idiosyncratic that it was unreadable to me in its original form. I hope
this never happens in the Apache Cassandra project.
I’ll close with a quote from an excellent article written by Colin
Breck, an engineer who works on large-scale data systems:
Colin wrote:
I don’t want to live in a world where you use AI to summarize
something important into unreadable text, and then I use AI in an attempt
to decipher it. I want to hear you, imperfections and all. I want your
interpretation of aesthetics, beauty, quality, relationship, time. I want
to know how you feel. I want you to cut through and tell me what really
matters.
Intentional writing will likely become more valuable. People who
write, and write to think, to think deeply and carefully, or to create, to
share, or to capture something important without explicitly expressing it
will continue to write and produce original work. The people who never were
writers will use AI to produce lots of text.
I hope that our culture can remain one of intentional writing and
intentional engineering. I enjoy reading the voice of the author in
comments, code, and tickets in Cassandra – the different ways we use
language based on where we grew up and how we learned English, the
translated idioms from our various backgrounds, and terse comments that
recognize the difference between code whose function is obvious and what
warrants genuine exposition. When I read code in Cassandra, it’s a delight
to recognize the author based on their writing style before flipping on
`git annotate` to reveal the origin.
I’d encourage folks to re-read the original proposal below. It is
very permissive. The guidance strikes me not just as reasonable, but
genuinely important to maintaining the health of the project.
– Scott
=====
Encouraged:
- Reviewing and otherwise validating human-authored patches before
submission
- Debugging, diagnosing etc
Permitted:
- Generating or modifying tests, scripts, tooling or any other
non-user facing changes
- Minor changes to human-authored patches that are carefully
reviewed by the author
Restricted:
- Core code changes made by LLM may only be proposed by contributors
with demonstrated expertise
- Must have produced similar patches in size, scope and area
unassisted and with minimal third-party guidance
- Core code changes made by LLM require an additional reviewer
- LLM review is not a substitute for human review, and must be used
only to augment a complete and independent human understanding of the patch.
Prohibited:
- All public prose must be human authored. This includes inline
comments, docs, posts to Jira etc.
All LLM generated changes MUST be disclosed:
- Outlined to any reviewer;
- Summarised in the commit message;
- Large blocks or files must be individually marked with some agreed
message like "created by <some AI>"
=====
On Sep 22, 2026, at 9:13 PM, Dinesh Joshi <[email protected]
<mailto:[email protected]>> wrote:
On Tue, Sep 22, 2026 at 3:28 AM Benedict <[email protected]
<mailto:[email protected]>> wrote:
Restricted:
- Core code changes made by LLM may only be proposed by
contributors with demonstrated expertise
- Must have produced similar patches in size, scope and area
unassisted and with minimal third-party guidance
I am -1 on this. This sounds like gate keeping attempt. It narrowly
limits the pool to a few people on the project that have historically
contributed to certain parts of the codebase. This policy will prohibit
skilled software engineers with domain expertise from proposing LLM
assisted changes simply because they have not contributed to the project.
This is unrealistic and a net negative for the project to attract talent
and grow our community.
- Core code changes made by LLM require an additional reviewer
Can you be more precise what is this in addition to? How many total
reviewers do you expect and what is the purpose of additional reviewer? and
why?
Taking a step back - what are you trying to solve here?
Dinesh

Reply via email to