[HN Gopher] Pinecone Open-Sources AWS Reference Architecture wit...
___________________________________________________________________
Pinecone Open-Sources AWS Reference Architecture with Pulumi
Author : zackproser
Score : 61 points
Date : 2023-11-27 17:26 UTC (5 hours ago)
(HTM) web link (www.pinecone.io)
(TXT) w3m dump (www.pinecone.io)
| manojlds wrote:
| Sounds silly to go to a vector DB vendor to get reference
| architecture for my app where it will only be a small part of my
| app.
| CharlesW wrote:
| I believe the point is that you can leverage the reference
| architecture for that small part of your app.
| zackproser wrote:
| Yes - this and to make it easier to understand how you'd use
| Pinecone at scale.
|
| We get a lot of feedback that our many open-source Jupyter
| notebooks (github.com/pinecone-io/examples) are very helpful
| for learning new techniques and understanding how patterns
| work - even for starting new applications.
|
| However, we're also often asked how to go from Notebook to
| prod. The Reference Architecture is a step in the direction
| of making this easier and more clear.
| whalesalad wrote:
| The massive (and contextually irrelevant) AI generated images
| every other paragraph remind me of setting line spacing to 2.2 in
| order to hit the minimum number of pages for your school essay.
| zackproser wrote:
| Thanks for the feedback!
| isoprophlex wrote:
| The massive paragraph heading size compared to the text
| itself doesn't do the readability a whole lot of good,
| either.
| cdchn wrote:
| Most of the text felt AI generated as well.
| zackproser wrote:
| Interesting! I can assure you none of the text was auto-
| generated. I've written before about how I don't like having
| text generated for me.
|
| I do use Grammarly for editing assistance when I'm writing
| something for work.
| _a_a_a_ wrote:
| Appreciated, but 16+ MB of images - there was no need given
| ~4K of text. That's a 4,000 times bloat. Please,
| professional devs don't need this. Just give us the info.
| zackproser wrote:
| I appreciate the feedback, but we may need to agree to
| disagree on this one :)
|
| As a professional dev, I enjoy articles with eye candy
| and AI-generated images, especially those with a pixel-
| art bent make me happy.
| nacs wrote:
| Except for the first image which is a human-made
| schematic, the rest of the images are filler and serve
| almost 0 purpose.
|
| At the very least they should have been sized down to 30%
| of their size and have the text flow around it instead of
| putting so many full-width images.
| paulddraper wrote:
| Don't forget to bump up those page margins.
| quickthrower2 wrote:
| They are the new cat images, borat jokes and rick rolling.
| ryoshu wrote:
| The isometric architecture diagram is hard to read as well.
| candiddevmike wrote:
| As someone interested in kicking the tires of VectorDBs, where
| does Pinecone rank? Is there one that will be "future proof"?
| federationfive wrote:
| Redis is probably better suited
| gk1 wrote:
| I work for Pinecone so my bias is obvious, but objectively:
| Pinecone is ranked as the most popular vector DB according to
| multiple sources [1][2][3][4], the best funded ($100M Series
| B), and is used in production by the likes of Shopify, Gong,
| Plaid, Zapier, Midjourney, CVS, and thousands of others.
|
| [1] https://retool.com/reports/state-of-ai-2023
|
| [2] https://state-of-llm.streamlit.app/#third
|
| [3] https://db-engines.com/en/ranking/vector+dbms
|
| [4] https://www.g2.com/categories/vector-database
| dmezzetti wrote:
| Couple relatively recent HN threads that give a good overview
| of the vector database landscape.
|
| https://news.ycombinator.com/item?id=36943318
|
| https://news.ycombinator.com/item?id=38416994
|
| https://news.ycombinator.com/item?id=38420554
| fzliu wrote:
| Regarding "future-proof-ness": we've been building production-
| grade vector search since 2018 and have a number of
| organizations running it at billion+ scale in production
| environments. It's all open source too.
|
| https://milvus.io
| redwood wrote:
| Anyone have good experience with Pulumi? The IaC space feels a
| bit crowded
| zackproser wrote:
| I've worked with Terraform extensively, including doing a lot
| of similar work (Reference Architectures, modules, AWS) over
| the past several years.
|
| This was my first time using Pulumi, and it was very much a
| delight. There's clearly been a lot of thought and effort put
| into the plugins and overall developer experience.
|
| For example, you can have Pulumi handle the Docker image builds
| for you - as well as the ECR repository logins and pushes. This
| means that your end users don't need to manually build images,
| log into ECR and push them - they just run `pulumi up` instead.
| rtuin wrote:
| My setup extends on this: the pulumi stack creates ECR repo,
| IAM user+access token, adds credentials and ECR details to
| GitHub actions secrets, and GitHub actions builds tags and
| pushes the images to ECR when a release is (automatically)
| tagged.
|
| Pulumi is (mostly) a bliss!
| Jemaclus wrote:
| I've used Pulumi for professional projects and personal
| projects, and I think it's fantastic. You can write your code
| in pretty much any major language (professionally, i use
| Python, personally I use Go), and you get the same output.
| Running the Pulumi program itself couldn't be simpler.
|
| My one critique is sort of weird in the sense that Pulumi
| supports way more than its public documentation says. I hope
| they up their documentation game in the future, because there
| are a lot of things that I had to go digging through source
| code to find out how to implement. But that's really a minor
| nitpick in the grand scheme of things. Big Pulumi fan here.
| quickthrower2 wrote:
| I agree! Looking at the source code is invaluable. Also if
| you use Azure, look at the generated ARM templates, for clues
| on how to set things up where the docs are short on details.
| It is nice they don't abstract over the ARM templates too
| much.
| cnunciato wrote:
| Pulumi docs engineer here. Really appreciate that feedback --
| what sort of thing did you have to go digging for? We're
| definitely investing here so I'd love to hear more about what
| you were looking for and didn't find (or what you discovered
| along the way that we could've made clearer). Thanks!
| rtuin wrote:
| There are more than a few resource types lacking import
| examples/details in the docs.
|
| Other than that, for most resources I use the docs are
| pretty good.
| quickthrower2 wrote:
| Yes Pulumi is nice. Having the power of Typescript types (if
| you choose Typescript) is amazing.
|
| The thing is I haven't used Terraform extensively so I can't
| compare.
|
| The main downsides I found are probably IaC downsides in
| general. Such as scenarios where you need to click-ops to
| repair because only some of the infra was deployed. I think the
| clouds need to make their stuff IaC friendly to avoid this
| though.
|
| My main fear is taking on that dependency on something newer.
| With Terraform if the company goes bust, the community could
| take over, and probably it would become an Apache project.
| Maybe Pulumi would too. It is not good enough for it to be open
| source. It needs to keep up with the daily changes in the
| cloud.
| zackproser wrote:
| Strong agree here - I am still keeping an open mind about
| TypeScript - using it for frontend / Vercel web apps and
| while maintaining open-source examples.
|
| I will say it was very nice getting the type hints for
| infrastructure while developing as well as hints about
| something being unused, etc.
|
| It overall worked better than I was probably expecting.
|
| Also agree with your points on the pain of IaC - it's
| terrific when it works but there's always those special
| "except this resource doesn't follow any of those rules"
| cases to make things painful and extra complex.
|
| I've also found that IaC can have a chilling effect for folks
| who haven't ever tried it before - it's a powerful
| abstraction tool - but at the end of the day you also need
| some level of understanding around what is happening under
| the hood - where to debug an error (which level), etc.
| kcorbitt wrote:
| For a previous startup a couple of years ago I tried to get our
| infra onboarded with terraform, hit a bunch of blockers,
| switched to Pulumi, and have used it for every project big or
| small since.
|
| Hits a really nice sweet spot for me -- yes, you want your
| infra definition to be declarative, but being able to write
| real code in a real language (Typescript in my case) that
| generates that declarative infra definition is exactly the
| right level of abstraction imho.
| next_xibalba wrote:
| Pinecone's aggressive marketing here on HN is getting really old.
| sabareesh wrote:
| Other day i was trying out Azure SQL for storing vector DB. After
| a POC i was able to get results on sub 1 second to get the
| results with 2 core serverless instance. It begs me to reconsider
| does dedicated Vector db really worth it ? Here is the article on
| how to store and query vector data
| https://devblogs.microsoft.com/azure-sql/vector-similarity-s...
| beoberha wrote:
| Ask yourself the same question about an OLAP system. Once
| you're at a large enough scale, it makes sense to use tools
| custom designed for your access patterns.
|
| I work on Azure SQL, but it's pretty clear that this is a cool
| demo. It may scale for some usages, but I wouldn't build a
| production system on it.
| sabareesh wrote:
| Well tried it with ~ 34 million records , which is good
| enough sample for our usecase
| bob1029 wrote:
| > Azure SQL
|
| > large enough scale
|
| You mean something like the Hyperscale service tier? This is
| what we are building our new products on top of.
| weberer wrote:
| >After a POC i was able to get results on sub 1 second to get
| the results with 2 core serverless instance
|
| Is that good or bad?
| sabareesh wrote:
| ~ 34 million embeddings , Seems this is a good enough sample
| for our use case.
| Dowwie wrote:
| Now THAT is developer advocacy-- a reference architecture and a
| complete video series. I'm intrigued by the fact that Typescript
| was chosen for this work. Are you finding that clients are
| rolling out ML architectures based on Typescript microservices?
| zackproser wrote:
| Thanks so much for your support
|
| We are finding that JavaScript is an often overlooked language
| for working in AI:
|
| https://www.pinecone.io/learn/javascript-ai/
|
| There's also a very nice synergy between using a language like
| JavaScript (many are familiar enough to read through and
| understand what's happening) and the type safety that
| TypeScript introduces.
|
| We are seeing a lot of interest in working with Typescript
| hence why many of our examples are TypeScript applications.
| dvfjsdhgfv wrote:
| Why is this on the main page? Just a few paragraphs of text with
| autogenerated images, and the contents feels like LLM-generated,
| too.
| zackproser wrote:
| Thanks for your question!
|
| The images were generated via DALL-E - so they're not
| technically "autogenerated". They were created using a
| generative model, however.
|
| The content itself definitely was not generated by an LLM -
| although that feels like a low-effort comment =/ What in
| particular felt LLM-generated to you?
|
| I think this is on the main page because being able to set 3
| environment variables and run `pulumi up` to get to a
| production-ready system in your own AWS account, without having
| to purchase anything, is a massive time-saver for anyone with a
| high-scale use cases that wants to use Pinecone's vector
| database.
___________________________________________________________________
(page generated 2023-11-27 23:01 UTC)