Faster – A key-value store for large state management | Hacker News Reader

Faster – A key-value store for large state management

microsoft.com

142 pointstim_sw34 comments

Does anyone remember MS Velocity / AppFabric Cache? No. Because it was crap and they canned it suddenly.

Really, if they are still supporting this in 3 years and it materialises into the promises I will take a look again.

I will never immediately adopt an MSFT platform component.

This is getting down-voted, but there's some truth here.

I didn't adopt .NET Core until v2, and happy I waited. Much stronger support from MSFT now, and a much more stable framework.

Exactly the other side of my point. Once it’s supoorted it’s good.

Insane performance on the most recent techempower benchmarks (usual proviso of all bencharks are meaningless etc etc.) they've caught Go with fasthttp already on one of them.

Not sure if it's good or bad, fast http is made by one guy and it's not official, aspcore is done by MS and dozen of engineers work on it.

both top the charts and are within 1% of the fastest on the chart so pretty good for the go dev and MS.

ASP.NET Core is a full-stack web framework and has about 100x more features and support. fasthttp is quick but it doesn't even implement the http spec completely.

For typical workloads that doesn’t make a whole load of difference. 10% difference in API meh. 90% of the latency and load is what goes on behind that and that’s your party not theirs.

This is a paper coming out of Microsoft Research. Not a product. Think of it more like Google's papers on Spanner, Andromeda, etc., except it's probably not in production use within Microsoft.

This isn't a commercial product or framework, it's Microsoft Research. They create many cutting-edge projects and do implement them internally in their own datacenters and services but they don't have the Google-like history of open-source clones based on the papers.

Velocity is the codename for the caching component of AppFabric, and AppFabric has evolved into Service Fabric, which now runs most of Microsoft's vast cloud services and has been recently open-sourced: https://github.com/Microsoft/service-fabric

Now you see here lies the problem. No one knew that who was using appfabric. We just got crapped on with this blog post and got told to move on and take our self hosted stuff to azure which at the time was a ball ache:

https://blogs.msdn.microsoft.com/appfabric/2015/04/02/micros...

The result of this is, to use the words I used at the time “well fuck you then” after we were told that this was the veritable jesus’s sandals of a product and the future of service oriented architecture at MSFT.

Now there’s a 180 on that post since, which we never checked up on because how would we know when we buggered off to memcache and foot the bill for the whole WCF and WWF rewrite debacle after 3.5. Then grumpily headed towards AWS and it was comfortable there. And now it’s open source and stuff (which judging by lesser parts of the ecosystem like SCVMM stuff on Linux) is probably going to be an abandoned wreck.

As always the marketing and roadmappery is crazy, disparate and impossible to track. We have no idea what direction MSFT is heading in, who is calling the shots and what’s going to happen next.

Which is the point. The maturity curve for MSFT software is a narrow window between two and five years. We can’t afford to rewrite our platform on that cycle.

And thus everyone leaves for greener pastures full of snakes, elephants, penguins and Bezos.

Fair point on MSR however.

I understand the situation, but have to disagree here. Let's face it, this industry is always in flux and things change, that's the only constant there is. However Microsoft is one of the best in the world at maintaining product support and backwards compatibility, often through decades. You can still run apps on Windows written in the 90s. You can even upgrade from DOS through Win 10 and still have a working operating system at the end. [1]

Yes, some small MS projects die, but overall you can still run almost all the old stuff, and interop and upgrade to the latest when you're ready. AppFabric extended support is through April 2022 [2], which is 4 years from now. It's possible you never got an update about it but regardless, they still support it. That being said, the cloud has changed everything and there are just more options now. In-memory distributed computing particularly has seen an incredible pace of progress with everything from Apache Ignite to Apache Spark so I think you would likely have more work to keep up with changes had you gone with another stack from the beginning.

1. https://www.youtube.com/watch?v=l60HHWWo9z4

2. https://blogs.msdn.microsoft.com/appfabric/2015/06/19/micros...

> To support failure recovery, FASTER incorporates a recovery strategy that can bring the system back to a recent consistent state at low cost

I'm at least intrigued, but this part is slightly worrying. "Recent consistent state" is unfortunately not "last good state", which would make it a non-starter for certain use cases.

> "Recent consistent state" is unfortunately not "last good state", which would make it a non-starter for certain use cases.

Is the issue "recent" or "consistent"? "Consistent" seems like a good metric for "good", so perhaps "how recent" is a problem you foresee?

The issue is “recent”, yes. I’m curious about “how recent”.

I wonder if this is a kinda of paper that is meant to be used under the hood of key/value in existing data bases vs a brand new data base?

I think you are spot on; this is an MS Research Blog post and not a product announcement. The most interesting thing in the post IMHO is the link to the paper. The Tango paper is another good read.

Yes, the article mentions that its siblings includes RocksDB. There are already outsized (north of 5TB) embedded databases geared towards high-IOPS storage (ie: NVMe) and this project serves to establish C#/.NET Core as a platform to build similarly-specced analytics products.

>First, we use a 50 : 50 Zipf workload, and plot throughput vs.RocksDB in Fig. 10. As expected, Faster slows down with limited memory because of increased random reads from SSD, but quickly reaches in-memory performance levels once the entire dataset fits in memory.

I dont mean to offend the authors in anyways. But can someone from facebook rocksDb team reproduce their results on rocksDB ? I am curious as to , why throughput remains constant, even though memory is increased from 5 -> 40GB.

Cool, implemented in C#.

What a unfortunate name. Good luck finding information when you need help with it.

Microsoft Faster is the first result.

Right bec it's in the news, wait two weeks, also the next 30 results are unrelated.

Also try searching something more realistic "FASTER lookup," "FASTER keystore," etc

I don't think microsoft.com will drop from #1 spot in search results.

I guess if all you ever need is the press release you're golden.

"FASTER access error" "FASTER login" "FASTER keys"

Not great.

So true, I hate when google does this too, what's wrong with using original names?

Not that this coulda been helped, but I spent like 2 hours yesterday trying to find info on the -filter param in powershell took me a very long time to find anything useful, and I still don't have the official doc on it.

Were you searching for things like "powershell -filter"? That would exclude every result containing "filter" [1]

[1] https://support.google.com/websearch/answer/2466433?hl=en

Yeah. I imagine everyone will call it Microsoft FASTER or MS Faster in Google land.

I believe this is just a Microsoft Research project. Based on the wording of the article I suspect they'll be using it within Microsoft products.

github or it doesn't exist

What happens when Faster deadlocks? It starves.

I want to play with it. Waiting for them to put the source on github. Too soon? ;)

Considering Microsoft is (and has been for some time) one of Github's most prolific users, I'd say if it's going to be open sourced, we should expect it soon.

Reply on news.ycombinator.com