Hacker News Reader

Comment by Pxtl | Hacker News Reader

I've seen many people describe frameworks like this - you know, first you have the slow back-end event-driven master database that you don't query live against, then you've got eventual-consistency flows against the various data-warehouses and data-stores and partitioned sharded databases in useful query-friendly layouts that you actually read live from... and I never see it clearly explained: how do you read a change back to the user literally just after they made the change? How do you say "other views eventual-consistency is fine but for this view of this bit of info we need it updated now".

This write-up is very detailed but I couldn't find that explanation.

View on HN

The short answer is write-through cache.

You write the update directly to the cache closest to the user and into the eventually consistent queue.

We did this at reddit. When you make a comment the HTML is rendered and put straight into the cache, and the raw text is put into the queue to go into the database. Same with votes. I suspect they do this client side now, which is now the closest cache to the user, but back then it was the server cache.

DynamoDB’s DAX cache espouses the same approach.

I have to say in my ~12 years as an active Redditor I can’t recall a time where I saw any real state issues, even with rapidly changing votes, etc. Bravo!? Now that we’re beyond the days of molten servers, I have to say its overall reliability in the face of massive spiky traffic is quite a feat.

Really? I see this all the time even now.

This explains so many bugs I came across on Reddit. I guess it works, but man I dislike this implementation.

Rama should bundle a write-through cache! Another in-memory JVM cluster thingamabob (Apache Ignite) used to propose write-through caching as it's primary selling point: https://ignite.apache.org/use-cases/in-memory-cache.html#:~:....

Or, maybe their pitch is that the streaming bits are so fast, you can just await the downstream commit of some write to a depot and it'll be as fast as a normal SQL UPDATE.

Rama is extremely fast, as you can see for yourself by playing with our Mastodon instance.

It’s fast until it’s not. Making a post and then hitting reload and not seeing it can be very jarring for the user. Definitely something to think about.

What do you mean? Every post I do shows up instantly.

Reloading the page from scratch can be slow due to Soapbox doing a lot of stuff asynchronously from scratch (Soapbox is the open-source Mastodon interface that we're using to serve the frontend). https://soapbox.pub/

I think the concern is will this still be true if Mastodon reaches Twitter scale?

Rama is scalable. So as your usage grows, you add resources to keep up. Scaling a Rama module is a trivial one-line command at the terminal.

Rama's built-in telemetry provides the information you need to know when it's time to scale.

is there a way to guarantee reading your own writes from a client perspective?

Yes. Depot appends by default don't return success until colocated streaming topologies have completed processing the data. So this is one way to coordinate the frontend with changes on the backend.

Within an ETL, when the computations you do on PStates are colocated with them, you always read your own writes.

It makes sense, but wouldn’t the write be slow? Especially when you have many streaming pipelines.

That's part of designing Rama applications. Acking is only coordinated with colocated stream topologies – stream topologies consuming that depot from another module don't add any latency.

Internally Rama does a lot of dynamic auto-batching for both depot appends and stream ETLs to amortize the cost of things like replication. So additional colocated stream topologies don't necessarily add much cost (though that depends on how complex the topology is, of course).

In Nathan Marz's (the article author) book, Big Data, he describes this and calls it the Speed Layer. I haven't fully finished the article yet, but the components it's describing seem to be equivalent to what he calls the Batch Layer and the Serving Layer in his book.

But I'm kind of getting the impression this works without any speed layer and is expected to be fast enough as-is.

Rama codifies and integrates the concepts I described in my book, with the high level model being: indexes = function(data) and query = function(indexes). These correspond to "depots" (data) , "ETLs" (functions), "PStates" (indexes), and "queries" (functions).

Rama is not batch-based. That is, PStates are not materialized by recomputing from scratch. They're incrementally updated either with stream or microbatch processing. But PStates can be recomputed from the source data on depots if needed.

So the idea is that you could do

1. send event data to depot

2. trigger localized ETLs (or put it high-priority in queue) to recalculate just the impacted data into relevant PStates

3. await completion of aforementioned ETLs

4. run query from updated PStates

Maybe too heavy for an upvote, but very appropriate for a an important transaction like a purchase.

Forgive me if I’m misunderstanding things, but this seems quite similar to what Materialize and ReadySet do, but like “as a library”, because Rama doesn’t use a “separate” layer for the storage stuff. Is that correct-ish?

I imagine you get some UUID back from your write, and effectively "block" until you see it committed to the event stream. The intent of such a system is certainly for the read-after-write latency to be not much longer than a traditional RDBMS. (This is roughly what the RDBMS is doing under the hood anyway.) Probably you can isolate latency-critical paths so they don't get stuck behind big stream processing jobs.

The advantage of the overall architecture is that nearly all application functionality (for something like a social network) can tolerate much higher latency than an RDBMS, so you really want to have architectural building blocks that let you actually use this headroom.

You can hack it and optimistically render the data you know about because your client created it - on the frontend, at no additional cost.

This is usually what I do. Don't even want to wait for an HTTP roundtrip for some of these, e.g. "liking" a post should fill in the heart icon or whatever instantly.

One famous example of this going to far: Mac Mail app used to play a whoosh sound when your email is actually sent. They changed it to whoosh instantly no matter what. Given how often an email might fail to send or get delayed, this meant an actually useful indication of "great, your thing was sent, you can close your laptop now" was rendered useless.

> Don't even want to wait for an HTTP roundtrip for some of these, e.g. "liking" a post should fill in the heart icon or whatever instantly.

HN does this, and on slow days, about half of my upvotes don't go through.

Messaging apps often have a checkmark to indicate the message actually went to the server, and maybe another checkmark to indicate it was received on the other end. Maybe HN needs an icon indicating that your vote went through.

Make the arrows grey to indicate the click registered, make them disappear to indicate the server successfully registered the vote?

Yeah, it's easy enough that I was able to do it in the web inspector in a minute (artificial 1s network delay added): https://s11.gifyu.com/images/ScPMI.gif

yea but does hn have any client side js?

Yeah, a very small amount so that clicking the upvote button does not need to reload the whole page

Can confirm, HN relies on client-side JS for voting and collapsing, but view/post/edit/delete don't need it.

makes sense, thanks for the clarification yall

Upvoting does work javascript disabled, it upvotes and reloads the page. It approximately keeps your position by jumping the message anchor.

I know it works with js disabled as well. But with js enabled they send just the upvote instead of reloading the whole page. Saving network traffic.

Unfortunately the js used is very minimal. And if the upvote fails, there is no graphical indication that it failed to upvote.

You actually check your list of upvoted comments?

No, I just notice it when I come back to the thread later in the day and a bunch of comments I know I upvoted are back to normal.

Yeah definitely, these ideas always sound very appealing to me, in theory -- I almost wonder why nobody has built it before

e.g. they mention "event sourcing" and "materialized views" in the post -- sounds good

But I thought I heard from a few people who were like "we ripped event sourcing" out of our codebase and so forth

And yeah your question is an obvious good one, and the Reddit answer of "write through cache" ... is less than satisfying to me

I FREQUENTLY have the problem where I reload the page and Reddit shows me stale data. It's SUPER buggy.

---

Anyway I definitely look forward to hearing people try this and what their longer term impressions are !

I basically want to know what the tradeoffs are -- it sounds good, but there are always tradeoffs

So is the tradeoff "eventual consistency" ? What are the other tradeoffs?

Hilariously, I went to edit the above comment, and HN was overloaded. Then it served me three or four 500's, AND it served me stale data in between

I was pissed off that I would have to type my comment again, but actually it did save it, and refreshing worked.

From what I understand Hacker News is architected more in-memory, on one big box ... Perhaps similar to the event sourcing model

(not knocking hacker news -- it's generally a very fast site, MUCH better than Reddit. Just that scaling beyond a single machine is difficult and full of gotchas )

StackOverflow used to run on a single (very beefy) machine also for a long time — databases make efficient use of vertical scaling, horizontal scaling is much harder.

Of course, specialist systems can often do much better.

When you step back and consider the incredible amount of manpower and resources that have been put into these applications, it's amazing how buggy these applications are. To put it simply, they're buggy because the underlying infrastructure and techniques used to build them are so complex that the implementation is beyond the realm of human understanding.

The way applications are built, and have been built since before I was born, is by combining together potentially dozens of narrow tools together: databases, computation systems, caches, monitoring tools, etc. There has never been a cohesive model capable of expressing arbitrary backends end-to-end, and every application built has to be twisted to fit onto the existing narrow pieces.

Rama is a lot more than just "event sourcing" and "materialized views". Those are two concepts at its foundation, but the real breakthrough is being that cohesive model capable of expressing diverse backends in their entirety. It took me more than five years of dedicated research to discover this model, and it was extremely difficult.

Yes, I 100% agree with you. I would like something like this to succeed, and agree the problem is real.

But what are the tradeoffs? There's nothing that comes with 100x benefit with no tradeoffs

(side note: I worked on Google Code for a short while in 2008, concurrent with Github's founding ... I think Github moved a lot faster in a large part because they weren't dealing with distributed systems at first -- they had a Rails app, a database, and RAID disks, and grew it from there. We had BigTable and perf reviews :-P )

Eventual consistency is probably one?

Can I specify that comment editing is correct and ACID, while likes/upvotes are eventually consistent? (No is a fine answer, these problems are hard)

I read through much of the doc, and don't see a mention of the word "consistency" at all, which seems like an oversight for something that is unifying what would be in a database with computation.

Rama is a much broader platform than a database, so the consistency semantics you get depend on how you use it. When using Rama, you're not mutating indexes directly like you do with a database, but adding source data that then gets materialized into any number of indexes.

You get read-after-write consistency for any PStates in a streaming ETL colocated with the depot you appended to. This is if you do the depot append with "full acking", which coordinates its response with the completion of colocated streaming ETLs. If you append at a lower level of acking, then you get eventual consistency on those PStates at the benefit of lower latency appends.

Microbatching is always eventually consistent as processing is asynchronous and uncoordinated to depot appends. Microbatching is higher thorughput than streaming and has simpler fault-tolerance semantics.

You'll be able to read a lot more about this when we release the docs next week.

I would like your newsletter.

OK, thanks for the response

I think many people are going to have problems programming with this consistency model, as they will with any that's different than a single machine. But that's basically "physics", so it's inevitable :)

But it seems like great work within the constraints -- look forward to learning more

I have indeed wondered why none of the cloud platforms have built more forward-looking tech like this -- instead it's copies of AWS and so forth

You should try using Facebook marketplace. It is so rickety. I have to get on a desktop to use it at all.

One strategy (somewhat common in lambda architectures) is to query both the long-term store and the in-flight operations, and blend the results. The in-flight stuff is both small and already in memory so it's pretty often trivially fast, even if blending the data is relatively complex.

That does limit you to operations/queries you can describe in this dual format, but pretty often that's fine. Or if you can relax read-after-write you can just ignore the in-flight stuff and read from the main store and then there are no (added) limitations.

You have the option to track the latest update time and, during the minute immediately following this update, direct all reads to come from the leader. Additionally, you could oversee the replication lag among followers and block queries on any follower that lags more than a minute behind the leader.

For the client, it's feasible to retain the timestamp of its most recent write. In this way, the system can ensure that the replica responsible for any reads related to that user incorporates updates at minimum up to that recorded timestamp. If a replica isn't adequately current, the read can either be managed by another replica or the query can wait until the replica catches up. The timestamp might take the form of a logical timestamp, signifying the order of writes (e.g., log sequence number), or it could be based on the actual system clock, where synchronized clocks become vital.

When your replicas are spread across multiple datacenters—whether for user proximity or enhanced availability—there's an added layer of complexity. Requests requiring the leader's involvement must be directed to the datacenter housing the leader.

It's called "read-after-write consistency". Write-through cache is one way, or just use a strongly consistent database, haha.

https://avikdas.com/2020/04/13/scalability-concepts-read-aft...

https://en.wikipedia.org/wiki/Cache_%28computing%29#Writing_...

https://cloud.google.com/blog/products/databases/why-you-sho...

Reply on news.ycombinator.com