FoundationDB by Apple; a great match for Elixir?

[andre1sk]

@OvermindDL1

That sound very slow, how do they pull that off?

I am not expert but copy pasting from the website

Transaction Processing
We often get asked how FoundationDB can achieve high performance for transaction processing over a scalable, distributed cluster. Here is a brief overview.

The transactional authority
Because FoundationDB supports concurrent operations by clients, it uses a distributed set of nodes working together as a transactional authority to detect conflicting updates. Transactions execute using optimistic concurrency control, so they don’t need to lock a key before updating it and unlock it afterward. Instead, when a transaction is submitted to the transactional authority, it will be rejected if there has been a conflicting transaction, requiring the client to retry the rejected transaction.

Transaction servers
To maintain high performance, the transactional authority is implemented by a number of individual transaction servers, each of which manages a portion of the incoming stream of transactions. FoundationDB’s design decomposes transaction processing into its individual functions and scales them independently. The separate functions are:

batching incoming transactions;
checking transaction conflicts;
logging transactions;
durably storing the data.
Of these functions, many people intuitively focus on transaction conflict checking stage as a potential bottleneck. Fortunately, it turns out that this function is scalable. FoundationDB uses a sophisticated data-parallel and multithreaded algorithm to optimize conflict-checking so that it requires only a small percentage of the system’s total work. This optimization allows a few transaction servers to keep up with a large cluster of storage servers.

[OvermindDL1]

@andre1sk

Because FoundationDB supports concurrent operations by clients, it uses a distributed set of nodes working together as a transactional authority to detect conflicting updates. Transactions execute using optimistic concurrency control, so they don’t need to lock a key before updating it and unlock it afterward. Instead, when a transaction is submitted to the transactional authority, it will be rejected if there has been a conflicting transaction, requiring the client to retry the rejected transaction.

This sounds like a bottleneck? Otherwise it sounds similar to the CAS instruction in assembly, take the ‘last’ state and if if the same then run these instructions else bail out…

@andre1sk

This optimization allows a few transaction servers to keep up with a large cluster of storage servers."

Hmm, yeah I wonder how they get around this bottleneck. Plus what if these servers go down? In Riak as long as a suitable number of servers are running then the thing is up, and even if worst-case certain specific nodes went down then only some data becomes inaccessible but the rest is still accessible. Hmm…

[andre1sk]

@OvermindDL1

Hmm, yeah I wonder how they get around this bottleneck. Plus what if these servers go down? In Riak as long as a suitable number of servers are running then the thing is up, and even if worst-case certain specific nodes went down then only some data becomes inaccessible but the rest is still accessible. Hmm…

I am guessing that that role is assigned dynamically using paxos. Again if you watch the video about how they are testing the thing it will become clear that they account for much harsher scenarios.

[cmkarlsson]

@andre1sk

FoundationDB uses a sophisticated data-parallel and multithreaded algorithm to optimize conflict-checking

I search for “data-parallell multithreaded algorithm foundationdb” but only came to the same documentation or this forum.

One paper did come up: “A mutlithreaded parallell implementation of a dynamic programming algorithm for sequence comparison”. But that seems to have to do with DNA sequencing.

To me this sounds like marketing speak ala MongoDB. Perhaps I am a bit too sceptical because promises made by other product-teams in the past but I’d rather be given a reference to the academic research behind this algorithm, and known flaws and weaknesses, than being told it is sophisticated.

[OvermindDL1]

@cmkarlsson

To me this sounds like marketing speak ala MongoDB. Perhaps I am a bit too sceptical because promises made by other product-teams in the past but I’d rather be given a reference to the academic research behind this algorithm, and known flaws and weaknesses, than being told it is sophisticated.

Yeah this is how I’m leaning and how it sounds to me.

[mischov]

Setting aside that marketing speak is generally marketing speak, I have watched the video that keeps being suggested in this thread and it’s very good and enough to make me less skeptical than normal.

If you check the hn thread for the open source release of Foundation, you can see that a number of people are curious about about conflict checking/resolution, and were answered, and searching there for “conflict” might eventually turn up more useful information.

[andre1sk]

I’d watch the video I am not eloquent enough to describe and can’t do it justice. If you have even a cursory interest in distributed systems you will not waste your time it’s very little about foundationDB and pretty much about using deterministic simulation for testing distributed systems. There is 0 reason for it to be marketing it’s not a product for sale or to be licensed it’s a technology Apple acquired in 2015 for internal use and recently open sourced a part of it.

[brazyl]

It really does sound very cool, but I must confess I just about choked when I saw that it requires both Mono AND a JDK

[joaoevangelista]

Isn’t it for compiling the drivers? It’s a monorepo afaik. Prettt

[brazyl]

João you’re probably right. I haven’t installed it yet or checked more deeply.

[brightball]

I thought it was strange that the Foundation announcement came shortly after the big announcement for CockroachDB. The two live in similar spaces.

[andre1sk]

Hmm CockroachDB has being around for a while not sure it’s related also FoundationDB has being used in production with multi petabyte datasets and largest instances CockroachDB is posting tests for are in few 100GB scale which is kindah pointless. CockroachDB has a good cross DC story and FoundationDB at least the opensourced parts are not really covering that use case well.