honestly the demo did the same thing to me, looked great then i tried to hammer it. what kind of load were you pushing when it dropped off, was it concurrent jobs or just sustained?

was gonna agree with op then i tried batching. takes it back. its fine

came in here ready to dunk on it too. but i went back and read the load test thread they posted and yeah, turns out i was hitting the wrong endpoint. retested clean, im walking back the hate. its fine actually

又一个画饼大师。demo好看有什么用，上量就拉胯。先把压测数据放出来再吹

works for me. no complaints

i was firmly in the skeptic camp last month, posted some salty stuff about it elsewhere. but ive been running a small pipeline on it for two weeks now and it hasnt fallen over once. i think the early builds were genuinely rough and people are still mad about that. credit where its due, the recent updates fixed a lot

still think the single job path is slow tho. batch saves it but not everyone can batch

wait what region are you on though? i heard the eu node is way slower than the us one rn

nah blockmachine is THE play rn 🔥 you guys are testing it wrong. batch your jobs and watch it FLY. ran 2k overnight no sweat

been on it a couple weeks and its been smooth for me, no drama. think a lot of the early hate was from the rough launch builds

load test or it didnt happen. dashboard means nothing

i want to push back gently on the batch everything take. for a lot of real workloads you cant batch, the requests come in one at a time from users and they need an answer now. so when people say it cooks under batch theyre kind of dodging the question op asked. the interactive single request path is the one that matters for me and that one is genuinely middling right now

dumb question but where do you even see the throughput number people keep mentioning, i dont see it on my dashboard

ok i came in to back op up but reading the batch posts i actually went and retested and. yeah. i was wrong, it holds. consider me flipped, the warm batched path is genuinely good

批量稳的 跑了几天没掉过

idk the batch thing feels like a workaround not a feature. if i have to restructure my whole pipeline to make it perform thats a cost youre just not putting on the dashboard

blockmachine speedrun: get mad, batch it, become a believer, post 🚀. weve all done the loop

本来跟楼主一个想法的，觉得就是个PPT项目。结果昨晚自己跑了一波 真香了 😅 批量丢进去稳得很，收回一半之前说的话

whats the cold start number anyone actually measured it

whats the max batch size before it starts dropping

serious q, is the dashboard number live or is it cached from like three days ago. because every time i refresh it says the exact same throughput and that smells

tried the batch trick everyone keeps pushing and yeah it helped but my single calls are still all over the place. so half convinced i guess

yeah blockmachine has been solid for me. ran a node on it for a few weeks, no surprise downtime, support actually replied. low key one of the better hosting options right now

认真问下，他们那个所谓的throughput是单节点还是全网加起来的？感觉数字水分很大

did anyone get the batch endpoint to work or is it just me getting 429s every other call lol

genuine question for the people running it longer than a day, what does cold start look like for you? thats the part that always kills me with these. the warm path can be great but if every fresh job eats two seconds of spinup it doesnt matter for my use case

问一下批量那个接口在哪开！🔥 找了半天没找到入口，是要白名单吗

想问下批量模式下成本怎么算的，是按job数还是按算力时长？我比较关心这块，怕跑起来烧钱

i spent the first half of this thread nodding along with the haters, im not gonna lie. then i actually sat down and instrumented my own runs instead of trusting the dashboard or the vibes here. with batching on and a sane retry policy the steady state is solid for me. i was wrong, the early skepticism was mostly me extrapolating from one bad afternoon

fun fact the warm path is like 80ms and the cold start is closer to 1.9s so when people say its fast and others say its slow theyre both right, theyre just measuring different things. anyway the answer to this whole thread is a p99 histogram and we have produced exactly zero of those

批量确实能打，我这边挂了一周还算稳。楼上吹的不算太夸张这次

at this rate the blockmachine docs are just gonna say step one: batch it. step two: stop complaining lol

im less sold than i was tbh. ran it again over the weekend and the tail latencies got ugly past a certain point. not saying its bad but the marketing oversells the steady state

我用下来感觉没楼里吹的那么神。前期空载确实快，但是一旦并发上去延迟波动很大，p99根本看不了。可能我场景比较特殊，但是这种稳定性对生产环境来说是硬伤，光看demo是看不出来的

still not convinced honestly. everyone defending it keeps moving the goalposts to the happy path. the single request numbers are the ones i care about and nobody wants to talk about those

im with op honestly. ive seen too many of these slick demos that fall apart in week three. ill believe the steady state when someones been on it a month and still happy, not after one good overnight run

for what its worth i had the opposite week. pinned it at a few hundred jobs and it held fine. wonder if its a region thing or which endpoint youre hitting

the demo is a sales tool. of course it looks good. post p99 or its noise

40 replies and not one graph. classic

love that the resolution to op being mad is everyone telling him he held the controller wrong. and somehow it works. ok blockmachine you win this round i guess

the batch defense only works if your workload can batch. mine cant. so for me op is right and yall are kind of talking past him

BATCH GANG stand up 🚀 ran a fat queue last night and it didnt even flinch. this is the most fun ive had with a subnet service in months ngl

i think the real benchmark here is how many times we can argue about benchmarks before anyone actually posts a graph. were like 40 replies deep and the only number ive seen is the 2k overnight guy and im pretty sure he was dreaming

kinda new here, whats the actual difference between this and just running the subnet yourself? is blockmachine a wrapper or its own thing

team batch vs team copium and im here for the DRAMA 🍿 just post the graph already someone

ah yes the classic, its fast you just have to use it in the one specific way that makes it fast. love that for us

yeah batching fixed it for me too, was about to rage quit then someone in discord told me to bump the batch size and it just worked

BATCH IT and it absolutely cooks 🚀 you guys hitting it one job at a time and then crying about latency is wild to me. read the docs

this thread is just two camps. team batch and team i refuse to read the docs. its beautiful really

lol the skeptic op walking it back two replies later was not on my bingo card but yeah, it does hold up if you batch. agree with the batch crowd

this thread has officially produced more philosophy about benchmarks than benchmarks. someone frame it

whats your retry config set to? mine eats 429s clean now

anyone know if theres a free tier or do i have to load up tao just to test