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.
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.
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.
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?
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?
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.
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.
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.
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.
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".
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 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.
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.
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.
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.
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.
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 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".
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.
> 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).
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.
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.
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.
Push-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.
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.
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.
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.
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.
Especially since in most of the cases where it's not-totally-insane to use, the right solution is the classic distributed database which came first, where the ledger is among predefined/controlled node-membership... as opposed to a bloated mass of workarounds-upon-workarounds to make it survive being ungovernable.
I've seen some boosters pivot to saying "private blockchain", but that's a vapid marketing-lie, a contradiction like selling "single-user Twitter" when really it's just a blog.
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.
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.
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:
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!
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....
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?
I reached out through email :)
Just one correction. My spec doesn't force /scroll/ URL's, just proposes it as a convention.
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.
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/
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 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.
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
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.
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.
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.
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.
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.
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 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".
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.
> 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).
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.
The proposed feed solution looks a lot like the CouchDB replication protocol, as another object in a convergent evolution space to consider.
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.
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.
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.
Push-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?
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.
> 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.
> 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.
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.
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 the classic distributed database which came first, where the ledger is among predefined/controlled node-membership... as opposed to a bloated mass of workarounds-upon-workarounds to make it survive being ungovernable.
I've seen some boosters pivot to saying "private blockchain", but that's a vapid marketing-lie, a contradiction like selling "single-user Twitter" when really it's just a blog.