The Valley of Webhooks | Hacker News Reader

The Valley of Webhooks

weli.dev

247 pointsweli102 comments

This is a nice writeup of the problems in using Webhooks for State Synchronization. I also noticed that the proposed solution is a pseudo IETF-style draft protocol called SCROLL... that happens to be remarkably similar to an actual IETF draft I am bringing to IETF 127 this November called "Braid-HTTP Subscriptions."

Both drafts request a subscription with a GET plus a header:

    Scroll Request:
      GET /scroll/feed/customers
      Prefer: stream
    
    Braid Request:
      GET /customers
      Subscribe:

In both systems, the GET leaves its response open to stream events. SCROLL responds with application/x-ndjson. Braid subscriptions are a 209 Multiresponse, with content-type application/http-history. This lets them support more than just JSON. You can send updates to the state of CSV, or PNGs, XML, HTML, plain text, or any media type.

The author noted that it's hard to get adoption. Well, the reason that Webhooks are so common is that they are bog-standard HTTP. For this to get adopted, we need to put it into bog-standard HTTP. So we need to go to the IETF, and and extend HTTP in a general way to support state synchronization. It should just work for any existing HTTP media type (not just JSON), and any resource/URL (not just special /scroll/* URLs), and any way of marking timestamps (not just the ordered strings proposed in SCROLL).

Then we can bake this stuff into HTTP, and thus into all our bog-standard libraries, utilities, and code, and you won't have to reimplement the same sync-logic-over-webhooks again, and again, and again.

Reach out if you're interested!

I reached out through email :)

Just one correction. My spec doesn't force /scroll/ URL's, just proposes it as a convention.

Is it accurate to say this is something like long polling except you continue to hold the connection open for subsequent updates? Does this mean a server potentially needs to hold open a very large number of connections (one per client) even if there are no updates?

And why formalize on HTTP rather than on a similar protocol over websockets?

And isn't that just Server-Sent Events (SSE)?

There’s also the Linked Data Event Streams (LDES) standard, which is a way of hosting a log as a set of linked http documents, fetched via polling and link navigation. Is there really anything more needed than a webhook for the notification and an LDES for the log?

https://semiceu.github.io/LinkedDataEventStreams/releases/1....

the solution is good but what tends to happen is webhook providers are not gonna work on such a solution - why - because it puts 'work' on them.

hell this is without the proposal for a new protocol - just a 'GET' stream or paginated one like the author said.

whereas with web hooks - a consumer has to do all the work - as the article above outlined.

Doesn't webhooks put even more work on them? You need to create a webhook dashboard, sometimes with multiple environments and administration endpoints, a way to set HMAC secrets, retriability and deliverability mechanisms, a local tunnel CLI for development, etc. With scroll if you already own a event log system the only work you need to do is exposing it to the client securely, same API keys as the rest of your existing REST api's.

I'm sorry, but I have to ask. bog-standard?

Never visited a standard bog?

cf. vanilla

It's a British colloquial term meaning basic / expected / ordinary

I had the exact same thing with the Quickbooks api recently. You cannot trust the responses or webhooks at all.

On create a user or invoice for example sometimes it will return an error, yet it actually created the entity. This means you have to check manually after creating everything to know if its created properly.

Then you have the issue that sometimes quickbooks takes a while to update, and locks the company file while it does some background magic. This means you cannot immediately do the existence check, and also sometimes the check errors or times out which essentially means you need to keep checking forever until you can properly reconcile your db against theirs. But with hundreds/thousands of transactions per minute this state is never reached. You perpetually live in a state of trying to catch up but never managing it.

When I brought it up with Quickbooks dev support their response was literally "Its your job to make sure things are created properly in our system".

How did we get to this place where we started putting up with systems that cannot ever be trusted?

That is the way enterprise software works as a system. It demands to be the central focus of everything. Workers want to route around these turbo productivity theater nonsense that could be replaced by a few K script that gates access to a text file and checks validity of appends. That can’t be allowed, so you need what is essentially whole poorly documented OSs to enable an economy of brokers to it, or the whole con would collapse.

> You cannot trust the responses or webhooks at all.

Well... yeah. I mean it's pretty obvious, no?

Here's some things that could go wrong regardless of what care the software tries to provide:

- The transaction completed on the backend cluster but the app instance died before if could create the response and after it committed the transaction.

- The transaction completed, the app instance transmitted a response, but the load-balancer/reverse-proxy in-between died before it could relay that response.

- Everything went well, but the ISP dropped some packets before it could get to you.

- Everything completed and the ISP stayed up, but on your end the response was flagged as malicious, or never made it through your load-balancer.

So, yeah. in general when you make an API request and get an error you have to check if the state was changed anyway, and if you aren't doing that you're doing it wrong anyway and cannot blame the system on the other side for returning errors.

Those are all network errors. You wouldn't just catch all errors and treat them the same. An application error should mean the transaction did not complete.

Of course none of that applies in this case, since it's about Quickbooks. If the quickbooks api says the sky is blue, you should double check just to make sure.

None of the errors you mention are what I am experiencing. Their system is creating objects but returning that it couldnt, not failing to send a response.

Some Apis are designed to always fail safe and send the correct response, or at least accept multiple messages without creating duplicates, so you can just keep firing the same message until you get the correct response. Not doing that is just lazy design.

Easy, distributed systems, coupled with folks that think they are the solution to what should be properly written modular applications in first place.

Instead they deep dive into distributed systems, without ever learning about them and all the issues that can arise.

There is a reason many CS degrees have two semesters full of distributed systems content, between networking approaches, architecture design and algorithms to make everything robust.

I much prefer cursor paginated API requests vs. webhooks. The obvious downside being that in order to not get 429'd you need a respectable poll frequency - meaning you lose reactivity to new events.

Thus I think webhooks still have a place - but as a simple "poke" that can be sent to the client to tell them something has changed - supplementing a default low frequency polling interval.

This gives us the best of both worlds:

1. No need to bother de-duping/retrying pokes - if you miss a webhook you will shortly recover anyway when you next poll. 2. No need for any local-specific tunnelling/tooling - the local app will work just fine with the default poll interval. 3. No need to keep a connection live for each client. 4. All the good stuff OP mentioned in his blog post.

Absolutely. I have been so livid at so many applications for not providing a decent CDC API (and dont forget deletes). Salesforce perhaps is the best out there. Imagine if every application exposed a standard CDC API, the world of integrations would be so much better.

Webhooks are fine, but a pollable API is a must have. The amount of hacks I had to do at work to workaround shitty APIs gives me nightmares.

The Gmail API works nicely like this. There's a history.list endpoint where you can see the recent history of message additions and removals, you can query for just the history that's taken place since a specific event's historyId, and you can subscribe to push notifications (that can be delivered by webhook) which just tell you when there are new history events, and you're expected to hit the history.list endpoint to see what's new. Some dropped push notifications aren't a big deal.

Yep, the poke pattern is additionally nice because it means my state reconciliation function is the same when running on cron interval vs event driven.

On the flip side, it helps to have endpoints which have a query param linking to some sort of resource update time stamp. That way you can query to only get those items changed since last poll.

With the proposed solution every consumer will have a persistent connection to the server irrespective of the frequency of events. This setup seems inefficient unless you have a very high volume of events coming in. Many CDN networks have a limit on how long a connection you can open. And data providers will not prefer serving persistent requests.

Problems listed are signatures, dedup, buffering, bootstrap, cron. Everything other than signatures and bootstrap, can be solved by having a counter in every webhook payload. It will increment each time. When you receive a webhook and the counter does not match, the consumer can fetch the missing data from the events API.

I agree with the author that providers simply saying "at least once delivery" is insufficient. they should have solutions that does not require an architecture diagram.

Bootstrap is better served with a bulk events API so you don't make one call per request. It can have an after/cursor pagination. Solutions that work for our internal Kafka might not be suited to work across services, over the internet.

Open TCP connections can also be wildly cheap and efficient - Apple Push Notifications (APNS) and Android's push systems maintain open TCP connections to just about every mobile device on this planet.

An open connection is just a bit of state on either end. The C10K problem has been solved for ages.

Anyone remember consuming Twitter hoses back in the day? Those were also long-lived persistent connections for efficiency reasons.

Not only this but the spec explictly allows provider-terminated connections followed by a 429 with a respected "Retry-After" if you want to definitely kick "slow" feeds. This combined with HTTP2/3 multiplexing makes holding feeds open pretty cheap.

Gerard mentions it super quickly, but another massive issue with webhooks generally is local development. Yes, you can use a tunnel, but that requires all engineers on a team to add their own tunnel urls. This causes even more issues when you use the platform as a source of truth, like for auth or payments. With WorkOS specifically, your whole team develops with one shared development sandbox. You run into issues when your local dev auth (in postgres) is not synced with the shared dev sandbox that WorkOS has since not all team members have their dev environments running at once. So yeah, then you use events API. But WorkOS only preserves the events API data for 90 days (and u have make 3 calls since its a max of 30 days per call). So then you load all the data with the state API first, then you start running the events API. It's a mess.

Tried to talk about this on X until the CEO of WorkOS wanted to bring it in private, then proceeded not to help at all. https://x.com/grinich/status/1913035839866835297?s=20

Webhooks are simple and ubiquitous, and that's both a weakness and a strength. It's also why they are used for a lot of things, even things they are not great for (state sync).

These weaknesses are why we[1] added FIFO endpoints, Polling Endpoints, and what we call "Svix Stream" as ways to do ordered state synchronization (each with its own tradeoffs). This lets people consume the events in the way that best fits their use-case. We are working on more things to make the state sync even easier. I'd love to hear about more challenges people are facing with webhooks, as we want to make these things better.

OP: I'd love to hear more about your thoughts there, and will send you an email in a moment.

P.S, if you're unfamiliar, please check out Standard Webhooks[2]. It's a spec we created to help with signature verification that has been adopted by OpenAI, Anthropic, Google, and many others. We are chipping at one webhook challenge at a time. :)

1: I'm the founder of Svix (mentioned in the post), we do webhooks infrastructure as a service.

2: https://www.standardwebhooks.com/

With webhooks, consumers get to asynchronously respond to updates from a provider. If no data has changed, a provider will not send any updates.

With SCROLL, consumers are responsible for choosing when to ask a provider for updates. Without a mechanism for knowing when data has changed, consumers will be forced to be pessimistic and poll providers for new data on some cadence.

I see two issues with the proposal: (1) SCROLL will lead to an increase in unnecessary network traffic for both the consumer and provider, and (2) because a consumer cannot know when data has changed, the lag between a consumer's local model and the provider's data model will be larger when with Webhooks.

Assuming you're not using the proposed streaming option, I suppose you could always send a webhook for that fact alone? In other words, an empty notification, with semantics of "something has probably changed, better poll the SCROLL if you aren't already".

Doesn't doing that just reinvent eTags on hypertext resources from the RESTful wars and XML Web Services days 20 years ago?

1. Long Poll the cursor to pull down the latest events

2. Trigger a long-poll even if in exponential backoff because they shot you a webhook saying 'eTag changed!'

If your API is just wrapping Kafka, it can long-poll

Webhooks are a painful problem. To clarify, Stripe's events API definitely ships a cursor and polling it has been the method preferred by large consumers for a long time.

Stripe events API is one of the examples of how to do things properly. And SCROLL is just trying to create a common spec so that everyone offers a stripe-like event polling api.

Yea I had the same thought: "It sounds like the Stripe events API would fulfill his needs." But then you said

> Nobody serves this today. That’s the catch, and it’s also the point.

and I was left wondering what was missing.

The core tradeoff here is "Push vs. Pull". Webhooks are a model for pushing data to subscribers. A traditional API allows clients to request and pull data. The data flow is in the same direction, but control flow is opposite.

Pull-oriented models are much easier to reason about and should be preferred where possible (cybernetically they are a closed loop, vs. push models which could literally just be a barrage of UDP packets). But they do have a little bit of overhead which makes them the wrong tool for some cases, like live-streamed entertainment or massive telemetry flows which value performance (latency, throughput) over missing a few packets.

You mean pull-oriented models are easier to reason about, right?

Yes, thanks. Corrected.

Just treat webhooks as a hint.

Write your code to work as a "reconciler" that checks the state of the remote system and reconciles it with the local view of that system. Run reconciliation for the full state periodically using a scheduler, and then treat webhooks as a hint to run the reconciler immediately.

This way, you will have a robust system that can survive logical bugs and outages because you don't store the synchronization state per se.

In the case of Stripe, for example, have a process that polls every open checkout session every couple of minutes. A webhook then just triggers the run earlier. If you're worried about DDoS, have an exponential backoff for the poll period.

Theoretically polling doesn't scale, but in practice it works just fine.

You don't really need a new protocol. If you trust your consumers, you can just give them a paginated "/events" endpoint and have them poll that. The delay will be variable, but tunable, based on the required timeliness. If you want more prompt responses you do long-polling or a websocket.

The key, and only thing that matters, is that the cursor rides in your database, and is therefore transactionally consistent with the event. That's the whole magic trick.

We've done event streams like this at the bank I work at for years.

This is kind of beside the point but... Shouldn't the the evolutionary metaphor be the opposite of how the author's using it?

They go on about how we've settled in the trough, but the whole local optimum problem is about settling in an early peak, and being unwilling to cross a trough to get to a higher peak. The workarounds would be propping up the first peak, not evidence that we've settled in the trough.

Either way, the metaphor feels kinda forced IMO.

Seems like we are trying to solve 2 problems at the same time with SCROLL:

* Database sync of log events. * "Real" time updates.

Webhooks can already mostly handle the "real-time" event portion. Of course there are problems. as the article expounds on, but a lot of those won't magically get solved with other solutions either. Distributed real-time communication is hard. Webhooks are good enough for this purpose.

For the database of log events, personally I'd rather just have a SQLite DB I can yank whenever. Don't give me CSV or JSON or whatever I have to parse and manage, just give me a SQLite DB ready to go. I'd love you for it. I'll just take a whole fresh copy with everything thanks. Maybe you limit it to to the last X events, say 90 days or 365 days or whatever, depending on sizing of events, but just send it all every time I fetch and I'm happy enough. If I need to generate a delta to keep some other DB in sync, well that's my problem. Just give every row a stable identifier.

Yeah, been fighting variants of this problem as a data engineer where the upstream database would only let you see the current state of the table and did not provide you a changelog / cdc

So you have to

(a) repeatedly query incremental date ranges, with some overlap (b) deal with late arriving rows

(d) query each day for all primary keys and remove rows that had been deleted (but only if you have newer data, otherwise you delete old values but don't have the updates)

(e) Sweat about if this goes wrong on a large table, how can you rapidly determine what state is out of sync? (Idea: how do I "hash" this in a way that hints at where the problem is?)

(f) Deduplicate, as mentioned

Martin Kleppmann's "Designing Data-Intensive Applications" book has some detailed discussions of some of the challenges...

This is a great article and exactly mirrors my experience consuming webhooks. My team built almost the same workarounds as the author.

The solution my team eventually settled on to the dedup, buffering, and race problems was:

  * When a webhook came in, store a single copy of the payload (usually json) in some temporary storage and enqueue the id.
  * Any subsequent updates (while the id was enqueued) overwrote that payload and skipped the queue.

So, we were able to ensure we only updated the object once and with the latest state (since the webhook payload always held the entire object state).

We didn't have a great solution for the bootstrap problem though.

I do like the SCROLL proposal, but I wonder about the cost of keeping around log-structured data forever - every log structured DB I know about does compactions for this reason.

Compactions are supported and heavily recommended! https://github.com/welidev/scroll/issues/1

The end here reminds me of "The Log: Real-time data's unifying abstraction" [0], which has unfortunately had a bit of link-rot since 2013.

One complication in this approach involves access-windows: What if my system is only supposed to be seeing stuff that happened during two separate weeks in the year, because those are the spans when it was subscribed or authorized?

So the data-host would need to maintain a concept of "connection history" for other services, and also use that to filter/modify its real event stream, inserting artificial "initial state" roll-ups of events that happened in dark periods.

[0] https://news.ycombinator.com/item?id=6916557

That requirement seems ... uncommon. How many systems/industries have a frequent notion of transient/sliced views of history for their customers?

If that is a real requirement, it seems like it'd be easier to meet by giving customers a realtime-stream/log API whose history starts when they were most recently granted access, and providing them older historical events via a separate API of the classic "ask for a report and we'll get back to you within a day or two with an S3 presigned URL" variety, then synthesizing that huge historical report in batch code that's aware of the subtleties of the customer's visibility windows.

Perhaps less a requirement and more a kind of consistency with prior limitations? For example:

1. I have events in a Calendar service.

2. I want to authorize Reminder service to see upcoming events, so that it can send reminders to attendees in a way the Calendar service does not directly support, e.g. SMS/WhatsApp.

3. With the necessary credentials/SSO, the Reminder service subscribes to Calendar and Calendar periodically POSTSs webhook updates. Reminder needs to recognize when an event is cancelled or rescheduled, so that can alter its reminders.

Do I want to give Reminder potential access to all events ever, or just ones active across the usage period? Meanwhile, the Reminder guys probably don't want to step through the whole Calendar-wide event stream to reconstruct which events haven't finally happened yet.

Fair, and thanks for explaining. Other than my original proposal of a simple webhook stream and a complicated offline backfill, all the good ways to do that with current tools probably involve building support for fine-grained read permissions (Reminder service's windowing) onto the producer (Calendar) side and emitting a tailored reminder stream.

If that challenge is indeed common, then there might be a market/demand/opportunity to implement ... I don't know what to call it, "Postgres RLS for Kafka" or something--a record-level security model for event streams, basically. Event compaction/deduplication would make this hard, though.

https://archive.ph/oIfml swerves the bitrot

"Nobody serves this today" except for every real estate listing service (https://www.reso.org/reso-web-api/)

I wonder if this is a CS problem somebody solved in 1954. Does somebody have the link to that paper?

(I'm not serious about 1954 in particular, I am about hoping somebody here knows the CS literature better than me)

The old timey systems modeled this problem using accounting 101.

When things change, you don’t immediately update the balance. Instead it is written to a transaction journal aka a log. The thing is this log is the source of truth. State or the balance is derived from the log.

You don’t send a continuous stream of logs. Instead it is batched and sent asynchronously. It is also applied asynchronously. It also records if the batch was successful or not.

If you have multiple systems sending their logs to a central server. No problem. The central server orders them all before applying the batches.

Every so often. The books are “closed”. Meaning the central server won’t accept any more journal entries for things that happened older than X dates.

I dunno about "old timey", I wrote a system that does exactly this like six months ago. Granted, it's a money tracking system.

It's always surprising to me how much of the real world runs on CSV and EDI files sent back and forth over SFTP.

I'm not an academic but I've worked through practical problems in this space for longer than I'd like to admit, often in ignorance of the literature. The most CS-y fundamental paper would probably be Lamport(1978) below, but the database papers are pretty fundamental on this topic in their own right.

Here are some pointers:

* Astrahan et al., "System R: Relational Approach to Database Management" (1976), oldest paper talking about logs in a database which is kinda what at its root this is recapitulating

* Jim Gray, "Notes on Data Base Operating Systems" (1978), covers logs with the goal of transaction management, undo, redo, the famous "two phase commmit" process for aligning state across a network boundary in two systems

* Leslie Lamport, "Time, Clocks, and the Ordering of Events in a Distributed System" (1978), not a database paper, more about state synchronization in general

* Bruce Lindsey, "Notes on Distributed Databases" (1979), standards for replicating data across multiple identical database nodes

* Jim Grey et al, "The Recovery Manager of the System R Data Manager" (1981), describes "write to the log first, then the database (WAL)" pattern

* Silberschatz, Stonebraker, Ullman: "Database Systems: Achievements and Opportunities"/"A Architecture for Heterogeneous Database Replication" (1980), discusses tricky bits about replaying logs in client database systems that work differently from the producer

* Leslie Lamport, "The Byzantine Generals Problem" (1982), on the math and needed consensus error handling when distributing state

For other later topics on state synchronization, read up on Paxos (Lamport, 1989/98), RAFT (2013), CRDTs (2011), etc.

The above approaches of "write-to-the-log-then-update-state" were applied to UNIX filesystem first in AIX 3.1 in 1990 and then adopted by other UNIX vendors and then in Linux ReiserFS/ext3/XFS(SGI) in 2001ish.

Lotus Notes took a different path in the 1990s to synchronizing state within documents which was not ACID-oriented like the above, but was more like the modern append-only-log with optimistic eventual consistency. Not sure about the best paper on this.

Then Martin Fowler popularized event sourcing in 2005 with Enterprise Application Architecture, and later described CQRS in 2011 and microservices. So "we" decoupled everything with Webhooks (2007) and Kafka (2011) and reinvented this problem. Oh, and did I mention Blockchain (2008), etc. Oh, did I mention Bitkeeper dvcs (1998) and git (2005) handling of distributed state?

Along the way databases (Snowflake, BigQuery, Iceberg, etc) started exploiting logs to show "older" state via features like "Time Travel" queries. Which is actually what Stonebreaker tried to do in the earliest versions of PostgreSQL but the first implementation in the 80s didn't perform well without good compaction support and on much more constrained resources.

Ask your local LLM for a good survey paper and it might be a bit easier to digest, but the above gives you some color and keywords.

This topic always surprises me. I do not understand the sequence of logic that leads people to build synchronisation mechanisms based only on webhooks.

Webhooks aren't at-least-once, nor at-most-once, nor are they guaranteed in-order. Some people build systems to make them more reliable, but if you really care about the data you need to think of a webhook delivery as best-effort, a bit like UDP.

That's before you get into all the extra complexities around these systems being owned by different people. For example, either or both system might have to roll back their database. Or either side might have a long-term bug in how they process webhooks, and now you have months of broken data.

My view is that the only reasonable thing is to start with the process that gets things back into sync if everything is broken. That almost certainly involves a poll or query of at least the upstream side, and maybe both sides.

I find that if you put a decent bit of engineering effort into that "disaster recovery" synchronisation, it can often act as the main or only synchronisation process for quite a lot of systems.

Stepping up from that, it's often useful to introduce webhooks as notifications only; that is, to provide a signal that some or all of the data is stale. You have to do a bit of consolidation, but this approach is usually enough to get completely reasonable latency for the kind of applications the author is describing.

Only if that wasn't enough for speed/scale reasons would I reach for a truly "push-driven" fast path. But you always have to be able to disaster recovery assuming the stream is wildly out of sync.

Some bits of the author's idea seem reasonable: certainly, I would love for there to be a standard protocol to request new data since some cursor or since some timestamp, ideally with some webhook notifications to give hints on when to poll.

The problem I have with the author's idea is that it is very strongly event-based, but the desired outcome isn't event-based. The desired outcome is almost always "the state over here looks like the state over there". Relying too strongly events ends up at the same kind of problem another level down: the "disaster recovery" script ends up wanting to compare the states anyway to figure out whether the events are broken.

Going fully event-sourced can work (although, I think, less often than advertised), but it really relies on everyone collectively agreeing on the same event stream being the source of truth. Once you start doing work across multiple organisations then that coordination is relatively rare.

What really surprises me is the variation in maturity on this topic. There seem to be people at all experience levels who are both doing this well and doing it badly. I have worked with people with decades of experience whose whole design just collapses if you ask "but what if X?" for some really banal values of X like "we have an outage for more than five minutes" or "we have to restore the DB to yesterday" or "someone, one time, accidentally merges a bug into master".

As an aside, I do find the obvious LLM-ness of the blog post and the proposal a bit disheartening. These are problems that require diligence and precision of thought. LLMs may be able to achieve those things, but that level of quality just isn't expressible in "Claudish".

Maybe the LLM has written maximum 50 words of the article by just being directed to switch things around and improve grammar and or internal consistency

"And here’s the absurd part: that history exists."

"It’s a jigsaw puzzle where the manufacturer had the original picture, cut it up, mailed me the pieces one at a time, lost a few in the post, mailed some twice, and printed nothing on the box."

"and that’s the entire problem: nothing announces a gap."

"None of this is any provider’s bug. Their webhooks work exactly as documented. The problem is what a webhook is: a notification, “something happened, here’s a POST about it.”"

Almost this entire section is clearly written by an LLM

I think my is-this-LLM alert is triggered not by mere use of phrases / constructions beloved of popular LLMs, but by things like unmotivated magpie-ish use of those phrases.

I didn't get that reading this (I didn't read the whole piece but I had read the parts you quote before reading your comment).

Often the LLM-beloved constructions are good usage in the right contexts.

Re-reading, I think the article is in a grey area, and I believe that there is a decent amount of human experience underpinning it. It certainly isn't 100% LLM, but it's a long way from 100% human.

But I wrote my commment after reading the article then the spec (https://welidev.github.io/scroll/), and so the spec was "top of mind".

The spec is just awash in LLM-isms. The cadence and rhetorical style are very Claudish. The visual style is basically "Claude's artifact plugin" (it may not be exactly that but it is an incredibly distinct signature). So the experience of reading the spec is very much an "AI slop" experience.

The reason I object to this is that the way these LLMs write is really well-tuned to gloss over small but critical details. And "small but critical details" are sort of the whole field of distributed systems.

This seems to be most true for Anthropic models (I am assuming there is some cultural defect in the way they give feedback), but it seems to be pretty universal, unless you give them some really strong stylistic anchor to a different style.

(As an aside, I sometimes wonder if this is part of the reason that LLMs seem from the outside to be succeeding disproportionately at mathematics: mathematics papers and mathematical notation may be a strong enough cultural force to override Anthropic's lack of taste and unlock the true power of the model).

I'm not saying there might not have been plenty of human guidance, but either way I don't think there's quite enough substance to this (based on everything I wrote in my comment) for this to feel like "a solution" either way.

Detecting LLM-written text is a critical skill these days. If you believe only 50 words of this was slop, then that's a skill you could improve.

> I find that if you put a decent bit of engineering effort into that "disaster recovery" synchronisation, it can often act as the main or only synchronisation process for quite a lot of systems.

Another benefit is that you get to exercise those disaster recovery mechanisms regularly as part of the normal functioning of the system, rather than a specialized path that is only rarely exercised (and thus may be broken when you need it the most).

I have seen it expressed here a few times, but these LLM blog post tech proposal things really give off the vibe (pun intended) of someone being gaslit by claude into thinking they are onto something really important or innovative. It feels really weird reading it. A bit like reading someone's somewhat self aggrandizing journal entries, or walking in on someone... Doin' stuff. I don't hate LLMs, they have made some aspects of my job easier for sure, but I think some people are still pretty off base on where and how it's acceptable to offload to an LLM.

Could be completely off base, but isn't this the equivalent of polling a Provider's commit log or distributed log? So accessing a third party's Apache Kafka?

Yep! And since your partner will be unlikely to build a new architecture on the request of one customer, you should 'do the Kafka thing' on your end, capture everything (double-sends and all), then smartly query your events, rather than try to program correct double-send-friendly db updates.

> Then the dedup table, because deliveries arrive twice and the docs cheerfully call this “at-least-once.”

Documenting unexpected or intermittent behavior: The easiest bug fixes of all.

The proposed “feed” solution is functionally indistinguishable from incremental reconciliation.

The article presents a good framing and is well-written, but doesn’t really propose anything new.

Yes, and I think that's good news. If what you need is log replication, having a standard means it's more likely providers offer an endpoint that conforms - and covers the common pitfalls (tombstones, what to do if cursors can't be sequential, etc) - because at scale any missed requirement can be a showstopper and send you back to emulating log replication by listening to out-of-order events.

> and is well-written

Nah, it's almost pure slop with just the em-dashes edited away.

Nice write-up of the webhook data consistency problem - I've been bitten by some of those Stripe issues myself. I agree that webhooks should come hand-in-hand with a "records updated since" endpoint that allows high-level discrepancy checks and API endpoints to pull full logs for any record. There needs to be some sort of reconciliation process that is both fully verifiable and not a firehose of data.

This is really a great article from implementation perspective and the challenges associated with it. The OP already captured the reasons. Even though the reason 1 has been mentioned it did not go deep into it and somehow focussed a lot more on 2

> 1. Trigger a side effect: send the receipt, start the build, ping the channel. > 2. Keep a copy of the provider’s data correct:

On high level

1. System can either be PUSH or PULL, webhooks are essentially push and towards the end the OP is exploring the possibility with PULL. The caveat is that OP already iterated the PUSH mechanisms thrice and is aware of all the hardships and is somehow hoping that PULL would solve them. Unfortunately the grass is same on other side too.

      a) The availability of the server can always be questionable in PULL mechanisms and its a lot of load on servers to support this kind of data at scale in bulk to multiple customers. You are essentially getting into database table scans. Its becomes a lot more costly with NOSQL databases. 
  
      b) The customer would end up making way too many calls to server even if data is not available or there would be additional latency when data was updated and when it was queried. This is one of the reasons why servers prefer to push instead of pull if they can find a listener available on other side. 

      c) CRLs (Certificate revocation lists) are good example which are available for PULL, same for all clients and yet rarely anyone does it correctly or does it at all even though its in security domain. In fact they are simple files on webservers in most of the implementations.

2. The primary use case for Webhook is for triggering the side effect and allowing the customers to choose if they want to subscribe for that event. A customer subscribing for everything even if its non-actionable should just treat it as logging data.

3. Logging data can and always have gaps, it should never be treated as source of truth. I might question the need for deduplication, usually there is a unique identifier and almost all databases support insert ignore kind of clause. Logging the event data just provides you with better availability and latency, the source of truth is still with the provider if a next step needs to happen.

4. If user cancelled the subscription in stripe and the event never arrived then its a system design issue or system availability issue on the client side. The complete data checksum or bulk imports at night are attempt to fix the problem in a hammerhead way . I understand it exists in lot of places, however it defeats the whole purpose.

> its a system design issue or system availability issue on the client side

If all events need to arrive, then the problem is not "notification" (which would be solved by webhooks) but "database replication": subscribe to new events, fetch the full snapshot, fetch the updates in range, have the monotonic value to establish the "range" in the first place. Reach the eventual consistency.

The proposed SCROLL handles half of these, which limits its use-cases.

i think a nice middle ground would be the "flip the arrow" but the webhook tells you when the cursor has moved and that's it so you don't hammer the origin server will polls or have N-time stale data.

I agree that this could be a good middle ground. This causes the webhook to be a push notification with the eventId/cursor, that then causes the consumer to pull the actual data (along with other recent events if desired).

I had this almost exact discussion today. Adyen (the payment provider) provides merchants with webhooks so they can update a local modal of payment and payment modification data (checkout).

But there is no way (for a merchant) to get the latest 'true' state as held by Adyen. So you better hope your data is exactly in sync with the notifications you got from the webhook (which it never exactly is, because there are so so many points of failures, and unlike what this author says, the docs aren't thát well presented to hold the same model as the PSP does. It is often close enough though, but you are constantly gardening your implementation, because the model also changes on their end with little information in the changelogs).

The "latest state" data exists though! If you open the customer portal it is presented to you without problem.

The proposed feed solution looks a lot like the CouchDB replication protocol, as another object in a convergent evolution space to consider.

I usually preach that a big blob of mutable state in the middle of your system (the db) is the problem, and events are the solution, and this seems like its most obvious case.

All your problems start when you try to modify your current state representation in response to hearing the USER_ADDED or USER_BLOCKED events. Just don't. Leave them as events. Any time you have a new stupid edge case (double send, out-of-order, add-then-delete-then-add), this becomes one new unit test, where you can soberly decide what it means, and update your read path to understand it.

If you bake the nonsense into the current db state every time a new event arrives, your first step in any debug or reasoning scenario is to unbake it: what events led to this mess? If you don't throw away the events, your debugging is done for you.

> Ask with a cursor and you resume where you left off.

How does the provider know what event the cursor you provided refers to?

Sounds like external state that needs to be managed ("cursor" -> timestamp)

That is completely fine. In the spec cursors just need to be lexicographically comparable. They could easily be timestamps if the provider chooses to. In many of my example they are just alphanumeric dictionaries.

The text ought to have been interesting, as the insight is real, but for god sake how can someone decently read the slop their LLM outputted as reply to their prompt “write a blog post about X” and just copy-paste as it is?!

Dude is not wrong ... but good luck convincing the Internet to switch to a sane system, when everyone already thinks web hooks are a "solved problem".

I can't imagine any reasonable engineer thinking webhooks are a solved problem. Imagine how much time and effort those provider must spent to offer high quality webhooks, to then still get customers complaining about missing packets and such. Webhooks are an operational nightmare.

This article resonates with similar struggles I've had at multiple companies. Its content is valuable. I wish the writing and the article itself had less of a slop odor.

That aside, what I don't understand (especially having worked on the side of the webhook sender, which is itself really tricky to get correct/performant/cheap) is why more companies which broadcast webhooks don't, say, provide direct access to Kafka topics, S3 buckets with ordered data objects landing, SQS queues, or any of the alternatives to those things.

"But it's irresponsible to expose an internal-use-only datastore directly to clients" goes one objection. But plenty of log-store systems have the notion of sharing a subpart of the log with a less-than-trusted external peer, so while exposing Kafka directly might be asking for the same kind of trouble as exposing your customer's SQL database for authenticated connection over the open internet (e.g. "we said you could issue reads, not that you could open/close TCP connections a million times a second! You just took out our message broker!"), exposing, say, an S3 bucket or Kinesis stream is much less risky because those systems have put some thought towards semi-trusted sharing.

"But everyone is used to getting HTTP webhooks and doesn't have the expertise to connect to something else"--that'd be true if, say, reading from a websocket or Postgres NOTIFY stream or Kafka topic or S3-change-notification stream were advanced techniques, but libraries around those things are so good nowadays that even the most web-tech-only low-skill developer can probably integrate with them with minimal hassle. Maybe it's just that a lot of shops literally only know how to run their code in a webserver, and have never deployed any other kind of application service/cronjob/queue worker? That seems unlikely to me, but I might be surprised.

"If we do something weird our competition will beat us on ease-of-use" goes another objection. But is it that hard given the libraries available? And can't you hedge back on the ease-of-use sell with "our data is fresher and more provably ordered and correct"?

I'm glad that SCROLL exists as a possible solution here. I'm just puzzled why more people haven't been using existing technologies to achieve this property.

Do most webhook senders literally not have a log store? Are they just firing webhooks in the middle of business event handlers and giving up synchronously if they can't be delievered?

Because if that's not the case (and I don't think it's the case), then it seems like the SCROLL API is ... basically just the Kafka consumer API. Or Kinesis. Or SQS. And so on.

We (Svix, webhooks infra) actually help our customers directly write to Kafka topics, S3 buckets, SQS, etc. and have for a few years now. There are definitely people that adopt that, but receivers as well prefer the simplicity of webhooks.

> receivers as well prefer the simplicity of webhooks.

Do you see receive-side customers as prepared to outright reject paying for vendors that only offer non-webhook event streams if there's a webhook-ful competitor available? Or is that preference more of the "well, it's easier to add a webhook route to our existing webapp than it is to run a stream consumer/cron/whatever, but neither of those two is cost- or effort-preventative" situation, where customers don't consider event delivery systems to be the main differentiating factor?

I know (from our customers) that some of their customers make buying decisions based on the availability of webhooks. I'm unaware of anyone doing it based on the availability of our other functionality (but it could be I just don't have the data).

Though the data I do have: how many people adopt these advanced endpoints in practice (as we offer these), and it's less than webhooks.

Interesting. There’s probably a market for the consumer side of that, then: really easy-to-integrate-with-common-app-frameworks-and-hosting-platforms runtimes that provide a polling/socketful event consumer and sidestep all the concerns that come with scaling webhooks on the consumer side. Your company looks like it helps with those concerns a lot! But if HTTP webhooks are the lingua franca, there’s only so much I can do as a consumer of them per TFA, you know?

> That seems unlikely to me, but I might be surprised.

I'm pretty sure you've hit the nail on the head. Especially now with LLM's you'd be surprised of how many people are running full on production apps that are coded solely by an agent, deployed with a mix of random providers (supabase + fly.io + whatever) and have no clue about what Kafka is. They are never going to ask for it because... they don't know they can, and why they should. They ask an agent "connect to stripe" and the agent codes your typical webhook ingestion mechanism, never proposes advocating for an event polling loop. Whenever there is a bug they will tell the agent to fix it and the agent will just write a webhook deduplication mechanism and that's it. You'd be surprised how little supposedly technical people care about the current state of technologies.

PS: Thanks for the feedback on the writing.

Can we ban AI slop posts?

If you have two or more services that need to agree about state, and you have some set of rules that govern what state changes are valid, and you don't want to mess around with any of this "what do we do when we miss and update vs when we get two of the same update" nonsense, and the services aren't in a position to query the same database, then you should really consider a permissioned blockchain. Consensus hard, but it's harder if you're not using tools that understand that what they're going for is consensus.

99% of the time (and all 3 times in the blog post), there is only one source of truth for any piece of data, and state transitions are completely arbitrary. Blockchain is almost always the wrong solution.

> Blockchain is almost always the wrong solution.

Especially since in most of the cases where it's not-totally-insane to use, the right solution is still the classic distributed database which already existed. In those, the ledger is kept among a predefined/controlled node-membership... as opposed to a bloated mass of workarounds and limitations to make it barely survive being ungovernable.

I've seen some boosters pivot to saying "private blockchain is good", but that's contradictory buzzword nonsense. It's like selling a blog as "single-user Twitter" or advertising a regular car as "user-controlled autonomous vehicle."

I'm not sure what a private blockchain is, but a permissioned blockchain is one where only certain parties have keys that allow them to write blocks. You end up with a message queue optimized to eliminate anything that would lead to the inconsistency nonsense that this article is talking about as soon as it is detected. It then becomes the writer's problem to retransmit in a way that doesn't cause a problem next time, rather than the reader's problem to recover.

The craziness comes in when you'll accept blocks from anybody willing to burn enough electricity to do so, or gamble enough tokens to do so, or whatever other artificial scarcity game people like to play. But if you're only planning to consume data from the eight other companies you do business with then there's no reason to bother with any of that, you can just hard code their public keys into your consensus protocol and you've sidestepped the nonsense.

What if the keys leak and need to be rotated?

If you're exchanging public keys to configure your node with in the first place, then presumably you have a trusted side channel for that kind of thing. It's equivalent to the webhook situation re: what if the domain name changes?

If you don't have such a channel then there's always social key recovery schemes, but I don't think there's a particular pattern that you can just adopt. You're now in participate-in-the-research mode.

If the hardcoded domain name changes, then the app needs to be recompiled with new domain name and that's it. The service is instantly restored and no data needs to be updated to reflect the change. In fact, the server side doesn't even need to be made aware there was any change. It just continues serving data as normal.

If you have a blockchain, and the blockchain requires a key to write to, and there's only one key, and the private key gets stolen by malicious email attachment and posted on darkweb - what now? Do you need a whole new blockchain? What happens to data from the old blockchain, how do you verify the new data is correct? Or do you keep the blockchain and use the leaked key to revoke the key and set a new one? How do we know the new key is legitimate?

I genuinely don't know how to recover in this situation and would really like for someone to walk me through the process.

That might be true for 99% of data, but I find that it's the remaining 1% of that data occupies the majority of the time. If you don't have a problem that is solved by a blockchain don't use one, but if you're saying things like this:

> I do not trust the copy I built, and I have no way to know when it’s wrong, so I will re-derive it from scratch every night, forever.

Then your life would probably be better if you just had to consume block-at-a-time and not the whole dataset every night. Better to be persistently five minutes behind, then to go all day not knowing whether you're wrong, with a brief moment of certainty each night.

Your nightly reconciliation job just ends up being an inside out version of the consensus protocol that you failed to enforce up front (which may be a necessary evil if you have no influence over the people who publish your data, but let's not let those people off the hook for failing to support incremental verification of sync).

Not 99% of data - 99% of applications. In 99% of applications, 100% of data is such that any piece of this 100% has only one authoritative source. Only in 1% of applications, the percent of data that doesn't have one authoritative source is less than 100%. So the absolute upper bound of when blockchain is even appropriate at all is 1% of applications.

> Then your life would probably be better if you just had to consume block-at-a-time and not the whole dataset every night.

Yes, exactly, that's the whole point of the article, the data should be an ordered stream, not asynchronous events. As long as it's an ordered stream, the author's life is peachy. Blockchain gives you an ordered stream, yes. But so does SCROLL. And if you choose SCROLL, you don't have to deal with the plethora of blockchain-specific problems.

Whenever you have a choice between SCROLL and a blockchain, you should always pick SCROLL and never a blockchain. Only if SCROLL won't work for your use case - for example, you actually need a consensus mechanism - you should consider a blockchain.

And no, blindly copying another database and overwriting every discrepancy with their version is NOT a consensus protocol! It's not meant to build a consensus! It's meant to copy data from authoritative source! There's no consensus to be had!

That is true, but I think it's a pattern we need to get away from. It's the source of everything that is wrong with the modern web.

- These single sources of truth become high value targets for corruption.

- Our apps become totally useless if they can't connect to whichever source we've hard-coded as authoritative.

- Even when connectivity is good and upstream sources are trustworthy, we end up creating unnecessary burdens on infrastructure for a connections that spans the globe from a single point to millions of them even when the data we're after is elsewhere on the LAN, ignored because it doesn't have the authoritative hostname.

It's much more fault tolerant to trust the data on the basis of its verifiable properties, not because of where you got it. That way, if the source becomes unavailable or untrustworthy, somebody more relevant can start being the leader (supposing you need a leader at all).

> you should always pick SCROLL and never a blockchain

The SCROLL RFC is very honest about how mature it is:

> draft-scroll-protocol-00 · request for comments · no implementations in the wild, one document

People have built some truly stupid things on a blockchain, and SCROLL does seem like a good idea for some cases, but I think you're going a bit far by suggesting that a protocol which has never been implemented is always preferable to one with dosens of implementations to chose from.

The real world operates on single sources of truth. The software merely models the real world - as it should. Saying we should move away from this pattern is synonymous with saying we should fundamentally change how are society works. Hint: if your preferred software solution only makes sense if we fundamentally change how society works, maybe it's not the best solution.

If Stripe says you didn't get the payment, then you didn't get the payment. End of story. The payment isn't magically going to appear in your wallet just because you verify some properties of some data.

If your only complaint against SCROLL is that an implementation doesn't exist yet, then how about implementing SCROLL instead of creating yet another blockchain? It's not much work anyway, SCROLL is like 1000 times simpler than even the simplest of blockchains.

I don't think society worked that way until the web made it possible for society to work that way. And it's not working particularly well. It's brittle, coercive, and it asks operators to have a hand in supporting it in any way to support it in every way, raising the challenge of participating to a level where only the largest companies can.

If we think we can fundamentally change it for the better, we should.

I'm not especially enamored of blockchains. I think that partition tolerance is more important than global consistency, so mostly it's CRDT's that I'm trying to apply. I just brought up blockchains because they seemed especially applicable to the problems that SCROLL is trying to solve.

If you have a single source and a reliable connection to it, SCROLL seems great, I just think we should be preparing fallbacks for when those assumptions need to change. Because the real world doesn't operate on single sources of truth. Pick anything that people talk about, and you'll find that they're not on the same page about which sources are trustworthy for that topic. The software world operates on single sources of truth not because the real world does, but because software has so far failed to be useful outside of that domain.

> I don't think society worked that way until the web made it possible for society to work that way.

I very much do think so - moreover, it's literally how it worked for decades, if not centuries. Before the internet, there was only one organization that knew whether you paid your taxes, the IRS. There was only one organization that knew your bank balance, the bank. There was only one organization that knew whether you have active water service, the water company. Everybody else - including yourself - had to either guess, or ask the single authoritative source.

Internet changed nothing about this. It only increased the number of queries and shortened response times.

Why is cryptographically verifying record sequentiality an important property here?

If I offer my customers a source of ordered records, the "trust" in that system is the fact that they pay me to make sure records are ordered. If I sell a fast or slow log database, approximately zero customers in the world care to verify ordering cryptographically.

Or by "blockchain" do you just mean .... records with sequential IDs? Because sequential, guaranteed IDs surface gappiness/idempotency a lot easier than Markov chains over cryptographic primitives.

That also doesn't address the other core problems in the article: the replication (or data retrieval/polling) protocol is a lot more complex than a blockchain's "I can verify and replicate the entire chain state from the beginning of time to you" single behavior. People want more specificity than that.

If your consumers are going to end up in this position:

> I do not trust the copy I built, and I have no way to know when it’s wrong, so I will re-derive it from scratch every night, forever.

Then guaranteed sequentiality means that they only have to verify each new block rather than fetch the whole thing every night.

Without it, you have this ever growing probability, which resets to 0 each night, that you unknowingly hold an invalid state. You might've acted on that state and so now when the nightly cleanup runs you have add code to go back out the consequences and instead apply the prosequences.

The complexity you think you're avoiding by not having a consensus protocol you're instead embracing as a data cleanup job, except instead of the same code everywhere, each reader has their own separate implementation.

That all goes away if you just don't process inbound data until you're sure nothing else is going to come along and invalidate it.

Sequential ID's work when there is only one writer and their implementation can be trusted but sometimes we get different events which both say they're number 12 and then we have to go call upstream and learn that so-and-so was on vacation and it won't happen again we promise. It takes days to resolve during which the potential of propagating bad state based on the lack of resolution continues to rise.

Ideally you can just avoid coordination delays entirely by keeping things monotonic and leaning on the CALM theorem but when that's not in the cards it's way better to put that delay on the writer's side, which is what blockchains do. Waiting for the consensus protocol to spit out a block before assuming that a write landed saves so many headaches on the reader side.

Reply on news.ycombinator.com