Latest / Agentic Conversations (formally mlops.community) / ML Flow vs Kubeflow 2022 // Byron Allen // Coffee Sessions #108

Transcript

ML Flow vs Kubeflow 2022 // Byron Allen // Coffee Sessions #108 · Agentic Conversations (formally mlops.community)

0:00
A buyer down and add your canno cortado.
0:02
I'm just pissed. I like my coffee with a lot of
0:05
ice and a lot of milk. Two years ago, Byron, you were
0:11
on here, we talked about the differences between Cube flow
0:15
and ml flow and we made a lot of cheese references.
0:18
It was a Cheesy episode. If I do say, so myself, we even
0:23
had a sponsor. We had a cheese sponsor.
0:26
That was the first sponsor ever for an ml Ops Community meet up.
0:31
And I can't remember who it was, but it was some random photo
0:35
that I found online. Hopefully, they don't sue us for
0:39
telling us. I did not think it was going to
0:41
get as much love and attention as it did.
0:43
But it turned out to blow up and it became like one of the most
0:47
watched videos and podcasts that we have ever had on the MLS
0:51
Community. It's been two years.
0:53
So I feel like it warrants a bit of an update, there's been
0:59
progress on both the cube Flow side and the ml Flow side.
1:03
And so I want to go into that. I also want to hear, I mean,
1:05
you've played around with it since then.
1:07
I want to hear your updated view point.
1:09
He's on it and just for fun. I've got a co-host Today.
1:13
George what's happening man? Pretty good.
1:18
Yeah. So we've got George who has been
1:20
killing it in the community. He's been absolutely awesome.
1:24
And I just figured, we got to get you on here, dude.
1:27
So I think about following me. And just so everybody knows
1:32
you've also played around with both ml flown Cube flow, and now
1:35
we're going to give our uncut, unfiltered opinions on the tools
1:40
now Byron. Maybe we can start with a bit of
1:44
an update on Just for those who didn't see part 1, what the hell
1:48
is Cube flow and what is ml flow?
1:52
Yeah, sure. So what is ml flow?
1:55
What is cuckoo below the seems late I think the big thing to
2:01
point out is that it is sort of comparison of apples and
2:04
oranges. However, they're both two tools
2:06
that came about in this kind of new round of considering, how do
2:12
we put ml models into production?
2:14
And they're certainly to that garnered.
2:17
A lot of attention, particularly around ranging, from like for 22
2:22
years ago and still to this day. And there was about two years
2:27
ago when we talked, I would argue that it was Probably Peak
2:30
hype for a lot of these. Like, if you think about like
2:34
Gartner hype cycle and how those work.
2:36
If you look at Google Trends is kind of interesting how that
2:39
overlaps with some of the trends for cupola and and I'm a flow,
2:43
but more for flow. But as to what they are and my
2:47
flow is a very succinct package, that is what I would describe.
2:51
As very empowering tool, it allows you primarily to do model
2:57
tracking and that's how I use it primarily I think a lot of
3:00
people kind of identify it as a tracking tool when I've done
3:07
architectural diagrams and described.
3:08
It also is like a an artifact store and I think it
3:12
facilitates. You know, the storing of three
3:15
main things you have metrics parameters and artifacts
3:18
artifacts including the model itself along with any kind of
3:21
other dependencies that might be used for regulated regulated
3:28
environment. So maybe you need some Mutation
3:30
that comes along with that says, certain that describes
3:33
attributes of what happened at the time of training.
3:37
So that's enough flow. It's it facilitates that through
3:40
a lightweight python package. There's Coupe flow, which really
3:44
was build more as an end-to-end platform, for machine learning
3:51
rml Ops. And it first came to, I think,
3:56
popularity my opinion, really, because it had Pretty defined
4:01
way of running ml pipelines. And that is probably one of the
4:06
core features that you would talk about when you're talking
4:08
about coop flow, but there's a lot more that kind of is
4:12
involved in it. And because it's a platform, and
4:14
the way it was developed, it basically was cobbling a lot of
4:18
other components together into one kind of unified platform,
4:23
but there's kind of some nuances to that.
4:25
I wouldn't say it's as unified as other platforms.
4:28
It's an open source tool. Tool it was two years ago kind
4:31
of in its early stages. It's not like a platform like
4:36
data brings words run by David. Brooks has a for-profit
4:40
organization and they kind of build them to host their own
4:44
software. Whereas Coupe, flow is an open
4:46
source tool. Yeah, hopefully, that's kind of
4:50
very succinctly, it's perfect. Good interior, and it's clear
4:54
concise and right to the point. Now, we get to get to the
4:57
opinions and all that good stuff, George I know you played
5:01
around with both of them. Maybe you can give us some
5:05
background on this. And I think one thing that I
5:07
loved when we first talked Byron was, when you told us about the
5:13
differences, you really like made.
5:15
Sure to make a point of ml flow is a pip install away and Cube
5:20
flow is like going to crash your computer a few times.
5:23
You didn't say that. You're very nice when you were
5:26
saying like how do you set up Cube flow?
5:29
Well, you got Gotta go through this and I'm for people who
5:33
listen to us a lot. They probably heard me say that
5:36
a few times but it's the joke that never gets old anyway.
5:40
George. What was your experience with
5:42
these man? I know that you played around
5:43
with both. Give me your your takes.
5:47
So I'm a flower I really enjoyed like I mean designed to be very
5:50
easy to use as a solar data scientist.
5:53
You can kind of get it, get it up and running integrate it with
5:56
a research project and do something useful with it.
5:59
I migrated to the Nano flow from tense aboard, which had had some
6:03
frustrations with. I think they were just, it was
6:05
quite difficult to compare different runs and you don't
6:07
have much control over, how you did it.
6:10
Or, at least, at least as I had it set up but setup, it wasn't
6:12
as intuitive as ml flow. I like that, it was quite like
6:17
simple and self contained. I get very frustrated by other
6:20
tools, which try to like, take on to bigger chunk of the pie
6:25
and then do know part of it like incredibly.
6:27
Well, you'd you'd want to you won't be able to chop and change
6:30
parts of your architecture. And I feel like tools, like you
6:35
flow, make that difficult to do whereas tools like NL flow, you
6:38
could swap that out for weights and biases, you could swap out
6:40
for tens of order. They've got their distinct
6:42
functionality and with things like that.
6:44
Makes it any easier for other tools that page was planning to
6:46
just integrate with all of them. They have a similar API Q flow,
6:52
if I'm very hard to set up and I'm still kind of trying to make
6:55
sure, I understand all the steps and that I can Integrate further
7:00
tooling with it and make good use of it but it's it's not an
7:04
easy thing for. So the data scientists do is
7:05
more of a cool, the devops team work with them and the vodka
7:10
together because you know what you want it to look like.
7:12
When it's all done, you know, the functionality that you need,
7:14
they know how to do it. And you just have to stay on
7:16
these very, very long calls where it's like, well, try
7:20
something crash the test cluster like have a bunch of problems
7:24
are hard to debug and repeat was this end?
7:28
You think this is because You just have to like the writing a
7:31
bunch of Gamble and you're not so fluent in.
7:34
Um, oh well, I mean, it's in my case it's also the fact that
7:41
when the point of these clusters is like a case of the tooling
7:45
that you have permissions and access to use as well.
7:48
Oh well I mean if you're if you're just working as a data
7:50
scientist, it within the team you will have the devops people
7:53
and it's like kind of dare tooling and you kind of do
7:55
things through them. So we have our Spinnaker UI You
7:59
can like work from the point manifest.
8:01
I'm starting to get it set up so you can use some commands
8:03
directly to that but like kind of restricted access setup.
8:08
And even then, when it comes to like lots of templated, ya know,
8:14
like you you're going to want to customize these setups normally
8:18
and and it's rarely simple or easy to do even like the
8:23
conflict mapping set ups. Always just 69 a lots of room
8:27
for little typos that up then hard.
8:29
Debug and hard-to-find aniline classic.
8:34
It's very different kettle of fish.
8:36
Yeah. And I guess that makes sense.
8:38
It's like you have to play in the devops field when you're
8:45
using Cube flow. Whereas like Byron was saying
8:47
it's like very much apples to oranges because ml flow is not
8:52
like that. So what makes it like ml flow?
8:54
So easy to throw in and get up and running so quickly and Don't
9:00
have to like talk to the devops team and then Cube flow you do.
9:03
Likewise Cube flow. So much heavier and ml flow is
9:07
not. Well, I just have two quick
9:09
comments, I find it. I haven't found any good way to
9:12
run like a lighter version of it, which is quite frustrating
9:15
because often, you know, that getting it fully up and running,
9:17
for the ready to use, is going to be a big piece of work and
9:20
you don't know if it's worth doing that however, much you
9:22
look through the docks. You don't know if it's worth
9:23
doing that. So if there's some stripped-back
9:25
version, which uses some lighter databases and like doesn't have
9:29
Every single component. It's really nice.
9:31
So you can set that up show your manager and say, I think this
9:34
will be really useful because it has these bits of functionality
9:37
that we're currently missing from our architecture.
9:39
We should go through the full effort towards the devops team.
9:41
Get it deployed so far. I haven't found that by might
9:45
just be not in the right bit of the dogs.
9:48
So I think that's a good point and I really kind of leads into
9:51
this idea of like developer experience.
9:53
And, you know, the first touch point for a developer experience
9:56
is documentation and I have never turned to ml flow
10:01
documentation and been disappointed.
10:05
I have definitely turned a coupla documentation and been
10:08
disappointed. So I mean that's the first touch
10:11
point, which usually doesn't like bode well for other aspects
10:14
of a tool if you're having a hard time with the
10:17
documentation, so I think it really does go back to developer
10:23
experience and it goes further than that.
10:25
I think if you look at him a flow as a tool it fundamentally
10:28
empowers. So George just described it a
10:31
scenario which is an empowering. It makes it an empowering tool.
10:34
He said, he could switch it out for other tools that are kind of
10:38
relevant related. That is that a very empowering
10:41
fact about ml flow. Whereas, if you have to sink in
10:45
a ton of time into understanding kubernetes as a data scientist,
10:49
which a lot of people are coming, That's from being data
10:51
scientist and not necessarily from being devops individuals
10:55
themselves. You know, they're going through
10:56
reaching out to their Dev Ops teams and organizations and
10:59
maybe they don't have cross-functional teams yet.
11:01
So the burden is on them, the day of that data, scientists to
11:03
go out and learn a little bit about kubernetes and then set
11:06
this up. If you're kind of in that
11:07
situation to then go and learn all this about, kubernetes to go
11:12
set up Coupe flow to go through that that process.
11:15
If the experience isn't superb, you know you're investing an
11:19
enormous amount Out of time to do something that's complicated
11:23
and you might not necessarily feel can be easily undone.
11:26
So you know, it's either kind of becomes like a sunk cost fallacy
11:30
in some ways. I think a lot of times, you
11:32
know, we see Investments and to certain Technologies and people
11:37
invest more in them because of Licensing or because they've
11:39
just built up a lot around it and it it's kind of like a
11:42
self-fulfilling prophecy that you keep going back to that tool
11:45
and investing in it because you're now reliant on it.
11:48
Regardless of how good it is is a tool But then, you know, you
11:53
have them air flow. On the other hand is empowering
11:56
and it enables a more flexible approach to ml Ops, which I
12:02
think is immediately visible when you start to use.
12:06
Yes, like that time to value is shortened and it's seen.
12:11
And, and I love that. You said that George.
12:13
Like, there's not a easy setup for it.
12:17
You gotta like really dig in and do a Of heavy lifting.
12:21
If you want to try and see that. And then what you're saying
12:24
Byron? It's like, yeah.
12:25
And you gotta go sift through the documentation, which is
12:29
leaves a lot to be desired. And so one thing it reminds me
12:33
of though like last week we just did this really deep dive on
12:37
Cube flow actually. And we had a cube flow fanatic
12:41
on the podcast which I love because I like to hear the
12:45
contrarian view, everybody likes to talk shit about Cube flow
12:49
including myself. And Ryan who was on last week,
12:52
just was like, yeah, man. I love Cube.
12:54
So I'm very bullish on it. And one thing he was saying is
12:57
that for him, he feels like you flow, is its most valuable, or
13:02
it has the most to offer. If you are a kubernetes first
13:07
shop, like if you're used to doing kubernetes everything,
13:11
then throw some cute flow on it because you're going to be much
13:15
happier than trying to like shoehorn in some kind of
13:19
whatever it. It is that you need an operator
13:21
to use or it's not like kubernetes - it's not built for
13:25
kubernetes and so you can get maybe 80% of the way there, but
13:29
as soon as you wanted like change something and tweak it,
13:32
customize it, you're not going to get there as easy as you
13:36
would with Cube though. One thing that I keep wondering
13:40
about though, when we look at all right, it's been two years
13:43
basically Byron. When I talk to you two years
13:46
ago, I had heard about Cube flow.
13:48
Probably like, A month before talking to you.
13:53
And so I knew very little and you were educating me and many
13:57
others that were on the call and then listen to it afterwards
14:01
about what exactly each did. Now I want to know, like what
14:06
has changed, what have you seen obviously ml flow is evolved
14:11
like a ton and it and we also had some of the lead mentee
14:15
maintainers of ml flow on a few weeks ago, and that was awesome.
14:20
Because Got to see their fundamentals and what they are
14:22
really like building for and this flexibility extendibility,
14:27
and then just like quickness to getting something up and
14:32
running. You don't need to get all of
14:33
them, L flow up and running. You can just get like one piece
14:35
and then it starts building out from there.
14:38
That was something one of their like core values.
14:40
And you see it and it's nice to see that you guys are talking
14:44
about that. So that's awesome.
14:46
But then like I also hear about we did a hackathon in the
14:49
community and and one of the things that we did was when pie
14:52
torch came out and you could serve models on ML flow with pie
14:56
torch. Well not when pie George came
14:58
out but when pie torch and ml flow integration came out.
15:02
So what gets me a little bit like confused is ml flow has
15:09
serving capabilities and then Cube flow has K serve and so but
15:15
then when I talk to people when I say alright would you consider
15:19
using cubed? Am L flow together.
15:21
And the guy like Ryan last week, he said, yeah, those are that's
15:26
a great combination but you use ml flow primarily for that
15:29
artifact store for those experiment tracking and you
15:31
don't do anything on the serving side.
15:34
And so I just like start to get confused as to where ml flow
15:38
stops and where Cube flow starts and like what you would use one
15:44
for and what you would use the other four and really like with
15:47
ML flow, is it good enough just to use ml flow on its own?
15:50
For this serving. Also and maybe Byron you can
15:53
like kick us off with this. Yeah, so I think one thing, you
15:57
said that does resonate with me, is this idea of that cupola is
16:01
good for kubernetes? First job.
16:03
I 100% agree with that. That makes sense.
16:06
It is a kubernetes based tool. And if you are founded and
16:12
dyed-in-the-wool kubernetes, I don't really think who flows
16:15
going to be hard for you. And you know what?
16:20
Our issues, their word that existed.
16:22
So a good example of this is there was a lack of
16:26
multi-tenancy for name space within some of the earlier
16:30
versions of coop flow. And you see this with Spotify,
16:34
they created a distribution or their their version of it, they
16:39
built around it to build around that pain point, right?
16:43
And so, but they were able to do that because they had that
16:45
kubernetes expertise and they had a central, it team that was
16:49
very familiar with the stand. A huge research.
16:51
They also had a mass. Did they had a big resource
16:54
base? They also have a really large
16:56
team to distribute to, so they had a there was sort of the
16:59
business justification to do it. Whereas that's that's not the
17:03
average job, you know, like most companies even when you look at
17:08
ml, I think today across the board really, you're not going
17:11
to be distributed like putting many models in production.
17:15
Unless you are your core product has ml around it or Your finance
17:22
company, which they tend to have more ml models in production
17:26
than others and then once you get to other verticals, it
17:29
really drops and at least that's what we've seen in our research.
17:33
So I think the Paradigm really kind of only matches some
17:39
companies, not all and yeah, so for others, you're going to have
17:44
to find another way that really fits in matches your business,
17:48
your data strategy, your infrastructure strategy, And
17:51
flows. It's not gonna really going to
17:53
be that for everyone. Yeah great point.
17:56
So now along the point of like serving have you messed around
18:00
with them L flow serving and George.
18:02
Maybe you have to like like the I think what we were doing when
18:07
we did the hackathon a year ago probably a little more than that
18:10
was. It was like torch serve and it
18:14
had ml flow capability or Mo flow integrated with torch
18:17
serve. I can't remember exactly what it
18:19
was now that it's been too long. Ang but have any of you played
18:22
with that unlike the serving with ML flow.
18:25
So I've used torsos actually I really liked.
18:29
It is quite quite simple, easy to deploy, like additional
18:33
models, easy to deploy the first Moto.
18:35
Everything's quite packet package quite neatly and cleanly
18:38
and very, very quick and easy to learn.
18:41
I haven't worked so much on the ml flow integration with it.
18:44
Does that track the is that on the on the monetary side then
18:48
what does NFL offer in that in that workflow That's what I was
18:52
hoping you could tell me Byron. So with so ml flow, you can
18:58
create any points with their mouth flow, and to the reality
19:03
for me. And most of my work is, I
19:05
wouldn't turn to ml flow as an endpoint, unless I was working
19:09
inside of a, an on-prem environment, because I would
19:14
just turn to the default cloud-based in point service
19:19
that exists on all A major major cloud provider.
19:22
So as your gcp and AWS all have some form of way to put a model
19:30
to an end point, and you can consume it from there.
19:33
So generally, when it comes to serving models, that's where we
19:37
turn to. In this case, most of our work
19:41
is completely on the cloud and on one of those Cloud providers
19:46
you can do you can again like serve a model using a muffler.
19:50
We've done that in the past and it works pretty well.
19:56
I've only used the rest API version of that.
19:58
Oh okay. I see.
20:01
Now we should probably talk a little bit about how It used to
20:08
be so cute. Flow used to have.
20:10
And as we mentioned, like you flow is, is a lot of different
20:13
pieces that are kind of strung together and built on top of
20:17
other pieces and I know like Cube flow pipelines.
20:21
A lot of people have there's been a trend to go towards using
20:25
our go workflows instead of cube flow pipelines, because maybe
20:28
it's easier. And it's what if, if you're in
20:32
devops, it's what you're used to and, and you feel more
20:36
comfortable with that. There's also what has happened
20:40
over the last two years? Is that KF serving, which was
20:45
one of the main pieces of cube though.
20:47
It graduated from Cube flow as they say or the cube flow
20:51
ecosystem and it became K Surf. And so what is interesting to me
20:58
about that is is that like what George was trying to mention
21:03
about some easy way to get a piece of like a The cube flow.
21:08
Maybe we can just like start using K serve instead of having
21:12
to stand up everything and I guess it'd have you messed
21:16
around with that like you know much about that Byron.
21:21
So when it comes to case serve, the thing that I've observed is
21:25
more kind of this cupola was this broad, kind of hard to
21:29
approach entity that was a collection of various different
21:34
components. It was a lot to bite off all at
21:38
once and they're certainly this momentum to try and make it one
21:43
unified platform for a while. I think there was some
21:47
recognition either that the market wasn't ready for that or
21:50
that, that was just too much in the initial instance.
21:53
There's also the the point that a lot of the serving aspect of
21:58
coop flow which now is K, serve, or is wrapped up into a
22:01
dedicated module casar. That was probably The more
22:06
successful part of coop flow next to a coupla pipelines in my
22:10
opinion. And you know, you see you saw a
22:14
lot of contributions to that from the likes of Google and
22:19
others but in particular you got it I think you got a really
22:21
named Selden. I think they really participate
22:24
in a lot in the development of that and they themselves are
22:28
with their own distribution and very successful.
22:30
Yeah, I was just going to mention Bloomberg to and shout
22:33
out to Dan's son who's like The heartbeat of that project and I
22:38
know he he put a lot into it but sorry for lying.
22:42
Don't let me cut you off, keep going with it.
22:45
Nah man. So I think to me it's not a
22:50
technical move, really? I didn't perceive it as that.
22:54
Anyways, it was kind of a practical one in terms of
22:59
positioning and moving towards, what's been most successful in
23:04
that space. so, if you look at it, it's yeah, I don't see it as
23:11
anything, more than that, really And I think what we also can
23:19
note since the last time we chatted was that there has been
23:27
the move from Cube flow and especially cute flow pipe lines.
23:31
Like now Google has vertex AI, which is basically where Cube
23:37
flow was lacking. It's like what Google tried to
23:40
make Barrett xai that end-to-end piece and a lot of it is built
23:44
on top of cube flow. And Byron.
23:47
I know you mess around a lot with gcp.
23:49
Do you have any insights on that?
23:51
And is it just like? All right, this is a great
23:53
option for those people who like Cube flow, but had a really bad
23:57
headaches or really bad experiences with it.
23:59
Now you can just have it be this managed service.
24:02
So I think I managed service of coop flow if you're not a
24:05
kubernetes, first shop is absolutely the way to go.
24:08
That's a great place to get started with Coop flow because
24:13
what it's doing is it's removing the infrastructure and set up.
24:16
It up through the managed service and then you can focus
24:19
on the actual data science and pipeline building aspect of the
24:23
work, which is where I think most people are really looking
24:27
for Value to be added by these kind of projects, you know.
24:32
They did like to avoid having to deal with infrastructure if they
24:36
can. So, in that sense, I think
24:38
vertex AI along the stage maker and others are really good
24:41
options. In that sense, specifically
24:43
vertex a I bring Then a lot of other services that are also not
24:48
founded in Coop flow as well. So there's a lot more than just
24:52
what could flow offered all in kind of one managed platform.
24:56
So I think that is the if anything you could say Google
25:02
kind of picked up the Baton where cupola was sort of leaving
25:06
off and I think made it more approachable for the greater
25:11
number of users. You know, George let's change.
25:16
Here's a little bit man and I want to hear about how you used
25:21
ml flow. So we've been talking we gave
25:23
Cube flow quite a bit of love there.
25:26
Now like let's get into the workflow that you have had when
25:31
it comes to ml flow and what did that look like?
25:34
Just walk us through that and and how it was for, you where
25:38
things were lacking or where things, you thought were
25:41
awesome. So I mean I was using it for
25:44
experimentation in active learning so it's all about
25:47
trying to find the right data points to most improve.
25:49
Your model was mostly sort of code written in pie, charts
25:54
lightning on the actual modeling side and it integrated really
25:57
smoothie really easily to have exactly the output I wanted to
26:00
see. So I needed both both training
26:04
runs at would just the conventional you want to debug
26:07
the the training process, see it?
26:09
See it Plateau. See how that compares on?
26:12
The validation of the test data, but then also wanting to plot
26:16
the sort of the best. Are you see, for each of those
26:20
models against different data set sizes.
26:22
So you when you're doing Active Learning experiments, you have
26:25
some active learning strategy, which selects the data points,
26:27
and you'll normally comparing that against random selection to
26:30
try to work out, how much of a gain using getting from that
26:33
data science, work in the first place.
26:36
So it's quite a lot of like just grouping the different
26:38
experiments together, making it very easy to then Sort of run
26:43
further analysis, on previous runs to, get that aggregated
26:45
graph out. I think, when I started to have
26:50
quite a lot of experience in there, got a little bit slow and
26:54
then maybe I think I remember getting a little bit frustrated
26:57
with the lack of flexibility in terms of the different visuals
27:00
that I created and so I kind of come from the data engineering
27:02
worlds like I'm used to True visualization tools, things like
27:07
superset, Tableau power, bi and so on.
27:09
And I'm kind of from the Data engineering solves, all of these
27:14
problems already going to get the data into a format where you
27:18
can use normal analytics tools and you can kind of you're free
27:20
to go wild, whatever you need to analyze a single model, great
27:24
off you go, you need to analyze like, how high parameters impact
27:28
performance? You've got everything you
27:29
already need in conventional tooling and probably already
27:33
designer ways that are more scalable.
27:35
It's meant for analytics of you know like consumer data and
27:39
things like that. So I guess that's kind of my My
27:42
overall frustration and findings.
27:45
Yeah, and I know that a few people when we were getting
27:48
ready for the chat with the mo flow crew, they were asking
27:53
about like when is ml flow going to have our back for example, or
27:59
when are they going to have the ability to to organize your runs
28:04
or not? You run your experiments in like
28:06
folders and that kind of stuff and it was very much like, yeah,
28:12
we've thought about it and we don't necessarily Really know
28:15
what the best way forward is but we want to talk to the community
28:19
and we want to figure out how to do it when it comes to like the
28:22
folders. But on the are back part, they
28:24
were like, well we do that with data bricks so that was kind of
28:29
like you want that. You got to pay for it and I'm
28:33
wondering if you felt if either of you felt like okay this is
28:38
great but it makes me want to jump to data bricks to get that
28:42
managed service and to get like some of these.
28:45
Enterprise features. The something about when an open
28:51
source project is designed in a way to lead you to the managed
28:54
service quite intentionally that restricts me so much that I want
28:57
to stay well, well away there's something about like, I
28:59
completely appreciate that, they have to find a way to monetize,
29:02
but some some companies and some technologies do that very, very
29:05
cleanly. And a way where you completely
29:06
understand why there's the managed service because a
29:08
certain area just is technically difficult like often when it's
29:12
the, in the sort of data. Engineering pipelining world,
29:17
all dogs, the prefect airflow, they all have their their
29:19
managed services and I completely understand that.
29:22
And it's a very, very clean division.
29:23
So they have their apis, which are fully open source, and you
29:27
never get frustrated. Because any feature that you see
29:30
and any feature that's available on the paper version, you get is
29:33
just that you have to do all of the very frustrating devops and
29:36
communities behind the scenes when there's an open source tool
29:39
designed to, like just with with gaps in the features.
29:43
And then you see that, open the open.
29:45
It on GitHub with like 50 likes of all the developers frustrated
29:48
by it and it's not going to go anywhere because it's never
29:51
going to be a priority because they want it to stay.
29:52
That way, that makes me want to just go for different tool.
29:56
I can definitely empathize with that kind of sentiment around
29:59
like that's that's obviously we've all experienced kind of
30:02
paid for tools or tools that are open source and kind of leaning
30:06
in that direction. I think for ML flow personally,
30:10
at least I didn't get it. Did so much of what I wanted it
30:14
to do. Without it being data breaks, it
30:17
didn't really convince me to go use data bricks.
30:20
I use data bricks separately in the past and yeah, I didn't.
30:24
I felt like they did a decent job of kind of separating it.
30:27
I know there's a couple tickets out there that I would have
30:30
liked to have seen done. And I'm pretty sure it was
30:32
because there was a date of bricks element at least, you
30:35
know, that's speculating a bit and the one that kind of stands
30:40
out is definitely authentication, you know, I've
30:43
always found it a little bit. That's kind of the one big hole
30:47
in my mind around ml flow is authentication because it's not
30:51
it's not immediately a platform in that way it's like a it comes
30:55
across as kind of a tool that's very convenient easy to use very
31:00
useful and helpful but there's more legwork there to make it
31:04
ready for adoption by broader team and that is really centered
31:08
around authentication and security by.
31:10
I feel it's quite a common common choice is this the same
31:13
with with Daxter which At all. I absolutely love the drill
31:18
expected to have some sort of way to easily integrate with
31:21
education and there's no tooling out there at all.
31:24
Almost yet designed to be difficult to do.
31:28
Sorry, we're going to say it goes just like it's the
31:33
Enterprise like classic Enterprise would need that.
31:37
So this is a perfect area for us to put like a cap on it and if
31:42
you want that, you gotta go and pay for it.
31:44
And so, I'm not sure though does data bricks have is that like,
31:49
if you get data breaks, then you get that indication.
31:52
I believe you would like, they should have authentication where
31:54
it's filtered down and from their platform and ends.
31:57
Yeah, ml flow in the rest of it all.
32:00
So I guess, you know, if you do, if you are finding yourself in
32:03
that position. Yeah, I mean I think this is you
32:07
can go to the platform but there's always this question
32:09
around TC 0 and this is true in all of the work that we do is,
32:13
you know, you can go out and buy Buy it, pay for it?
32:16
Or is it worth just building at your facility?
32:19
And I think sometimes it really is, you know, worth investing a
32:22
little bit of time to just build that apparatus that ensures
32:26
enough authentication and security that you need.
32:30
You know, that's those are kind of skills that you want to have
32:33
around anyways. So it's I don't think it hurts
32:37
to attempts to facilitate that. Yeah.
32:39
And go out and do that to kind of do I feel like there's
32:42
something that like, it's Done before.
32:46
There's got to be some project out there, right?
32:48
That somebody's done it and they open sourced it to the world.
32:52
So if if you have an add-on for Mo flow, tell us, we'll give you
32:57
a shout out because we would love to see that and I know
33:01
there's a managed ml flow service that was popping up in
33:05
the community of few months ago. And they were talking about how
33:09
they were ml flow. That was managed but it wasn't
33:12
data bricks. It was just someone that fork
33:15
time. Mel flow.
33:15
And so there is a lot of Buzz around the mo flow community and
33:22
also the cube flow community in general.
33:24
This next part, I wanted to dive into just those two communities
33:28
and Byron, you've mentioned the docks and how they look in the
33:34
two aspects like the ml flow docks and the cube flow talks.
33:38
But when we look at like community community support,
33:42
just Community activity and know.
33:45
Met a lot of people make their choices based on the strength of
33:50
communities that in open source project has behind it.
33:53
Like, what in your eyes does? The ml flow Community have
34:00
versus the cube flow community? And maybe Byron can you kick us
34:05
off and then we'd go to George. So I think the immediate first
34:09
thing is you have a huge background.
34:10
The form of data bricks for ML flow, and there's individuals
34:16
are driving that whereas it's a, the Coalition that is backing
34:21
Coupe. Flow is a little bit more
34:23
wide-ranging. There's a lot more people
34:25
involved, also the goal and the aspiration of good.
34:29
Flow is a far bigger goal and aspiration.
34:31
Again, we're talking about apples and oranges here, and
34:35
it's substantially. It's It's a bigger pie, it's
34:39
more to deal with and so I imagine the coordination and
34:42
that is more challenging as well.
34:44
So it's not as simple as it all flow in that sense and I think
34:49
that has you just kind of see that in the documentation and in
34:54
code, not only that but also like examples that they put
34:57
forward, there's some examples that don't work out of the box
35:00
in certain clouds. So this there's things like that
35:04
that make Make it more complicated.
35:08
The communities I guess themselves, there's aspects of
35:11
the communities. I'm not 100% aware of because
35:14
I'm speaking as to both tools as an end user or a consumer of the
35:19
tool, not as a participant of those communities.
35:22
So I kind of see it from the outside.
35:24
I mean, I see a lot more coming from Mo flow but I guess that's
35:27
because it's got the lower cost of adoption.
35:30
There's more users. So if you ping out a question,
35:32
you're more likely to get quick, quick answers on how you can
35:35
sort of move forward, how you can integrate it Project how you
35:38
can track a certain metric, Etc. I think this probably largely
35:42
based on the fact that people actually working on on the
35:45
deployment of Q flow after your in further between its people in
35:49
larger companies with great engineering resource and
35:52
probably more senior as well. So you get that a little bit
35:54
less help but maybe, maybe higher quality, I'm not sure.
35:58
Hey, I'm Vishnu, I'm a data scientist at first hand and I
36:01
definitely think that you should subscribe to the ml Ops coffee
36:05
sessions podcast, it's the best. Podcast out there to stay on top
36:08
of what Emma Labs actually is to talk to the truth, thought
36:12
leaders in the space. And oh, by the way, Demetrius is
36:15
absolutely hilarious. What a weird guy, you should
36:18
definitely subscribe to the podcast.
36:21
That's such a good point. I'd also add to this that the
36:25
because of the complexity and the way that the history of coop
36:27
flow has gone. And as we talked about, there's
36:31
sort of this movement of, yes, people is an open source
36:33
project, but it's really been picked up by Google in the form
36:36
of vertex. You also have a number of other
36:39
services that are putting this into their core product.
36:44
I think Rick do is a very obvious one, but then there's
36:47
like another like two, three of them trying to do something very
36:50
similar to this Selden. Yeah, Selden seldom very much
36:54
focus is on on model serving so they kind of have that that
36:57
Niche covered as well so you can definitely include them in that.
37:00
But in terms of like trying to offer like an indent platform,
37:03
its kind of segwayed into that. Hammock where people companies
37:09
are picking up the open-source tool in order to generate a
37:12
product. Whereas that doesn't seem to be
37:15
as often the case with ML flow, it actually does seem to be a
37:18
bit more of a community from what I can tell.
37:21
That's a huge point. It's like there are different
37:25
incentives in the cube flow community and different people
37:30
that are looking to use Cube flow in ways that are more
37:35
self-serving. In ml flow, I do see that like
37:40
George. I feel like that is more common.
37:42
What you were talking about? Where people are more apt to do
37:47
a toy project or write a blog post about how they're using ml
37:53
flow or how you can stand up this XYZ system using ml flow.
37:58
I've seen a lot more of those than I have of cube flow.
38:03
Like your here's my, here's how you can do XYZ with.
38:07
I'm Flo and that's probably goes back to the he's he's ability.
38:12
Like just stand up ML flow. It's a pip install away like we
38:16
say, but then in the ml flow Community you have the one big
38:23
player which is data bricks and then you have everyone else and
38:27
data bricks knows what they're doing and they're putting the
38:30
resources and energy into where they want to take ml flow and
38:35
it's clear at least From there, you don't have it being
38:40
splintered in so many ways. And then like you said Byron,
38:43
it's not as big of a project. So you don't have all these
38:49
areas where it could splinter in different ways.
38:53
So I think both of those are really valid points.
38:57
On that point actually Demetrius you mind if I share my screen
39:00
real quick, this might be of interest to you.
39:02
So I'm gonna you're going to see a flash of something that's
39:05
unrelated but there's a developer experience which I
39:08
think there's a lot more depth to how this interplays with
39:11
both. But that's actually not what I
39:12
wanted to share with you. I wanted to share this Google
39:16
Trends chart, but I think is quite oh yeah, I'm gonna start
39:19
to look at it. So we got like, green is stage
39:23
maker which should not really surprised.
39:25
A lot of people practicing in the Area.
39:27
That that's what kind of is coming up in Google changed for
39:31
those that are just listening. You can see that Sage maker just
39:35
like basically skyrockets up and is an arm and a leg above any of
39:41
them. So you have Sage maker and greed
39:45
that's arm and a leg above. Yeah.
39:46
And they, you know, they were some of the most.
39:48
It was one of the earliest ml ml Ops platforms that came out and
39:52
so there are fast-moving first adopter.
39:55
You see, not pretty soon. After that you see eval flow in
39:59
pipe flow kind of take flight. The interesting thing I see here
40:02
is that you see ml flow in yellow.
40:06
It kind of goes through this like dip of like, actually this
40:10
This is the month that we had. Our last talk.
40:12
I think it was July 20. 20 was like P type of both of these
40:19
tools and then there's kind of this, trough of dish.
40:21
This illusion allusion meant as they say.
40:24
But you see vertex H AI R, sorry, Coop flow in red, doesn't
40:29
really quite recover in the same way that ml flow does.
40:32
So, I think there's, I mean, arguably, you could say there's
40:35
if you are trying to read in between the lines here that it's
40:40
They're what it came out of that successfully, whereas Coupe flow
40:43
didn't as much, but equally at around the same time you see
40:47
vertex a I really pick up. So I think like to me this
40:50
really speaks to the point that you have that handoff between an
40:55
open source tool and the more managed service versions of it.
41:00
Mmm, yeah I remember when the ml Ops Community was just starting
41:05
and we had I can't remember when we were about about to get to a
41:10
thousand people in the community and the cube flow Community
41:13
already had like 2,500 and it was huge in my eyes.
41:19
And at some point in 2021, it went from like Cube flow
41:24
Community had 5000 and then the mooc community had 5500, it was
41:29
like, whoa, we've got more people than in the cute flow
41:31
Community, but then, it kind of felt like it got neglected in a
41:35
way, because a lot of the people that started Cube flow project
41:41
started going and doing different things and being
41:45
associated with different projects, or they were
41:48
interested in other new things. And so you it felt like it got
41:54
neglected and then I heard from people who would go to the town
41:57
halls in the cube flow like Community calls.
42:00
That it's like, there's not that many people that are showing up
42:03
anymore. It doesn't feel like there is
42:05
that energy behind it. Like it there once was and I
42:08
remember a lot of people. Has to come into the ml Ops
42:11
community and ask Cube flow specific questions.
42:15
Because in the cube flow Community, they couldn't get
42:17
answers. And that might be along the
42:19
lines of what you were talking about George is that a lot of
42:22
people that are using Cube flow in a really like productionize,
42:26
the environment are a bit more advanced, or they're at these
42:30
bigger companies, and they have much more resources.
42:32
So they're not going to be hanging out in the cube flow
42:37
slack as much as the people. We're just trying to stand it up
42:41
to experiment with it. So that being said, I mean, I
42:47
also love the fact that you put on there, the ml Ops term search
42:52
term and that one has been steadily growing.
42:55
So that's cool to see but we haven't reached Peak hype yet.
43:01
Yeah, I mean I think it was important to put that because it
43:03
kind of it showcases the trend the topic is still high, you
43:07
know. This isn't about this isn't as
43:09
much. There's a tooling aspect to our
43:11
broader conversation around. Amell yeah.
43:14
Such a good point, and I think that is like, probably the
43:18
biggest thing for me when you were mentioning at the
43:21
beginning, George about how Cube flow is difficult.
43:25
And it's heavy is because it's not only that you need to be
43:28
setting it up, but you need to go.
43:29
And you need to ask for access from devops Team.
43:33
You need to get like, more people involved in this
43:36
endeavor. It's no longer just like, hey
43:38
can I hack this thing together and show some value?
43:41
It's like, all right now I gotta have some time.
43:44
Time with the devops team and then we got a debug and now it's
43:47
not just one person that is on this experiment.
43:51
Its I'm roping in more and more people.
43:54
And that is a organizational, problem, more than it is a
43:58
tooling problem, right? Like, that is trying to figure
44:00
out this this from various stakeholders point of view.
44:05
So it's such an interesting piece that we have here now.
44:13
I don't know where we're going to go.
44:14
From here. Is there any like final
44:17
thoughts? You all have when it comes to
44:19
cube flow and ml flow anything that you would like to see from
44:24
either of these? We talked about ml flow, we
44:26
talked to like seeing a bit more of these authentication tube
44:33
flow. Maybe just I guess I'm not sure
44:38
what we want to see from Q fo some smaller easier, bite-sized
44:42
piece of cube flow would be Great.
44:45
Any last words. Maybe George.
44:47
You got something for us. It's just got out for a second,
44:51
but I think, I think from my perspective, lots of these,
44:54
like, specialized and I'll tools can be broken down into a
44:59
specialized modules on things, like the evaluation side which
45:02
and I'll flow does really well. But then some of the tools can
45:05
just be borrowed from data. Engineering world.
45:06
I feel like ml pipelines. That sort of problem is solved
45:09
quite well by things like Daxter prefect flow.
45:14
And so on and when it's when apis are clearly defined, it
45:18
gives you so much. It's empowering like Byron.
45:20
Said earlier, I completely agree with that phrase.
45:22
You a project is very intimidating and not enjoyable
45:25
if it's going to be hard to do a 180 on.
45:27
You want to be able to try something.
45:30
Even start to integrate all your all your projects with it and
45:32
then to be able to swap them out because you realize that you've
45:35
made a slight mistake and it's missing a feature in terms of
45:39
like developer experience, like nothing improves it, as much as
45:42
just sort of that. Very, very contained API and
45:50
yeah, for a small shop, like my company, things like, when
45:55
you're going for something like with the complexity of Q flow, I
45:57
feel like the cloud vendors, make more sense.
46:00
It's only if you're much much larger organization, where the
46:04
sort of managing your own own queue, for instance, might be
46:07
worth it. I think there is a like, who
46:10
floated, find a home in my mind through these different managed
46:13
services. And that's to be honest with
46:15
you. The way that I think a lot of
46:18
where, a lot of what we're doing in this in it and the broader
46:24
data space, a lot of things are moving towards the managed
46:27
services or manage versions of themselves or integrating ML.
46:31
And to databases, you know, there's all sorts of there's
46:34
always a movement towards trying to make something a little bit
46:37
more automated a little bit more convenient.
46:39
So I think the movement of coop flow into a managed service.
46:43
Space is a good. One.
46:45
And I think there's a place for it because that's really where
46:49
you're going to see empowering. The end-users, the data
46:52
scientist, the ml developer that's where they're going to
46:55
get empowered by a pipeline such as that is if they don't have to
47:00
manage the infrastructure themselves.
47:02
So I think it's not a bad thing that this is kind of the
47:06
direction of where kupo went. It just makes means of the open
47:10
source Community isn't as robust but it's hard when you have such
47:14
a Project. So last one that comes to mind
47:19
is how ideally would you use Cube flow and ml flow together.
47:25
So ml flow I'd describe it as an artifact store and the
47:30
alternative in Coop flow is not robust enough and I think Amell
47:37
flow. You can log out to amount of
47:39
flow from your coupla pipelines. That's a possibility.
47:44
Although I have heard someone describe it as hard to justify
47:47
because we'll why am I not? Why am I not just using ml flow?
47:51
I mean that's that's not entirely a good way of looking
47:53
at it, but that's sometimes the kind of thinking that Creeps in
47:56
when you try to mash them together.
47:58
Is that people will say, well, what's the point of coop float
48:01
supposed to do this end-to-end solution?
48:04
You know why, using a wide bring another tool into it.
48:08
So, I think that's sometimes a harder conversation for some
48:11
people, but I think it is doable logging out too.
48:15
That were other systems is possible and we've done that.
48:18
So I actually there's one engagement that my company had
48:22
been a part of where they were logging out not to kupo but to
48:26
just bi dashboards and that was one way that they that's how
48:29
they chose to track and perform observability in the space 33.
48:35
Well, this is awesome. This has been great guys.
48:39
I love it and hopefully this has been an update for everyone that
48:43
has been looking. For where we are now when it
48:46
comes to ml Flow versus Cube flow or if you're just getting
48:50
into the game and you're wondering what all these flows
48:54
are. I can think of like five
48:56
different throat flows off the top of my head, we've got air
48:59
flow, we've got cube flow, we've got meta flow, we've got ml flow
49:05
was another one that I'm missing.
49:06
There's all the flow brothers that are there and then there's
49:10
a bunch of stores to. So I did not coin this term, but
49:14
We could call it the flow brothers and the store sisters.
49:17
So you've got your metric store. You got your feature story, you
49:22
got your evaluations store, all different kinds of stores that
49:26
we can make a whole Mall out of all these stores.
49:29
Put them together and have your commercial shopping mall.
49:32
So that is it? I think we're going to end there
49:35
and this has been awesome. Thanks again, guys, thanks so
49:39
much. Thank you.
49:42
This is the bonus. We've got here.
49:43
Byron. We And also, we wanted to talk
49:46
about entity Centric stuff, right?
49:50
Yeah. So, I mean, this could almost
49:54
wonder if it's a separate podcast, but I'll tell you what
49:56
I think. So, I really love the podcast
49:58
about Maxime fisherman, and I hope I'm not butchering his name
50:02
and I apologize if I have, but he's a legend in the community
50:06
and a lot of what he was saying in the podcast.
50:09
Really resonated with me, especially, given recent work.
50:11
So he mentioned the idea of Entity eccentric modeling.
50:17
And so I mean entities, eccentric modeling has been
50:21
around. I would argue that if you want
50:23
to say something's entity Centric, we've actually had that
50:25
like Star schemas you usually have a table and that it has an
50:29
entity associated with it. So in that sense, it's entity
50:32
Centric, but if you start to look at this evolution in the
50:36
broader data stack towards not being constrained by storage or
50:41
CPU, in other words think. So, Like think bigquery, think
50:46
any kind of modern olap data warehouse, you can scale very
50:50
easily to a lot of data which means you're modeling, Dynamic,
50:53
can change. You don't need to have a star
50:55
schema where you do as many joints, you can start to
50:58
denormalize. And if you start from the
51:00
position of having a denormalized table, I think that
51:04
that's actually really handy because that means you can start
51:08
to do group buys across more rows with fewer joints.
51:13
And what that allows is is a lot less overhead.
51:18
When you're modeling is what I found and it fits well with the
51:22
recent experience, I've had using the looker semantic layer
51:25
and the semantic layer. There is a pretty thick,
51:28
semantic layer compared to what you might see in superset, which
51:32
I think has been described as a thin semantic layer.
51:35
Either way, I think the idea of having a semantic layer that
51:38
helps build out that data. Model is very handy in a
51:41
paradigm that I really enjoyed working with, and I can't help.
51:44
But think, well, why is why are we not doing the same?
51:49
Like having one semantic layer for both bi and ML and starting
51:54
from the position of denormalize bottles.
51:56
Because if you start from the position of a denormalized
51:59
model, then you start to work towards the aggregation is that
52:01
you may need for both, which are usually along some kind of
52:05
You're aggregating along product or your are getting along claims
52:11
if you're an insurance etc, etc, whatever it might be people
52:14
users so on and so forth. So I kind of wonder if there's
52:19
in our industry room for a semantic layer.
52:24
That caters not only for bi, which is what you see with the
52:28
semantic layer and superset and looker, but also caters for a
52:32
future store. So you're going to have to break
52:35
down exactly what you mean by semantic layer because I've also
52:38
heard this in the DBT community and I've heard that like the
52:44
people that I've talked to were not happy with the semantic
52:48
layer options that they had for machine learning.
52:52
And the way that I understand it and you this, why I would like
52:56
you to break it down is like the ability to know if I change any
53:01
data. What is going to happen
53:04
Downstream? Or who that data effects is not
53:08
clear to me and people were promoted proposing.
53:13
That a semantic layer would help you understand better, the users
53:17
and the consumers of your data and I know this is all.
53:20
So as we start to go a little bit towards like the data mesh
53:24
theories and when we start talking about consumers of data
53:28
but what is a semantic layer? Because then I tried to tell
53:31
people, oh yeah, this guy told me that a semantic layer would
53:34
be great for that, and They were like what the hell's a semantic
53:36
layer. So break it down for me, so So I
53:41
think just going back to that point about data mesh.
53:45
I think, did a mesh is a buzzword right now and I'm not
53:49
in this. I mean I see it as a list of
53:51
pain points and not necessarily a list of solutions to get to
53:56
the point where you have a solution though for those kind
53:58
of pain points. I think you do need something
54:00
with like a unified semantic layer but so that's kind of
54:06
starting off at like where we might be going to in the future.
54:09
Not where we're at today in terms.
54:11
What is a semantic layer today? As I know it?
54:14
Because semantic layers can be different.
54:15
This is really about how you how you define a semantic layer as
54:18
well. So let's take for example,
54:21
looker the semantic layer there. Is it looks like Jason a Json
54:26
file and it's basically a config file that describes your data
54:31
model. And when it's executed, what it
54:34
does, is it builds up the SQL to query your underlying database
54:39
and query the data that you need.
54:41
So instead of having everything kind of stored in a static table
54:45
at the final layer, which you would do often for some older bi
54:50
tools, where you needed to have increase performance, you would
54:53
build the data model to a point where everything's and that
54:55
layer and just clear the table directly.
54:57
But it's already aggregated like a highly aggregated values.
55:00
This isn't true for all bi tools.
55:02
There's some that use kind of different caching mechanisms in
55:05
between to support a better more performant experience too.
55:11
But in this case, what it's doing is it's using that it's
55:17
building out that SQL query through the semantic layer
55:21
serves. Basically as a view and you use
55:24
that to get your, grab your data and then there's some smarter
55:27
caching mechanisms that can be put in place to reuse the data
55:32
so that you don't have to recreate every single time.
55:35
So the semantics going back to what is a semantic layer?
55:38
It can kind of, it's up to the developer.
55:41
And one semantic. Layer does not equal, another
55:44
semantic layer for sure. It's there to generate code in
55:50
another language so that that is what a semantically.
55:53
Hmm, I see. And then you were talking about
55:57
a unified semantic layer like how in your eyes, does that
56:01
look? And then you said well and then
56:03
we can throw a feature store on top of it.
56:05
I kind of see how a feature story would fit into this.
56:08
If it was bad to use cases, I guess I understand.
56:11
And correctly. Yeah.
56:14
I'm I'm thinking that she's case feature store not online.
56:19
I think online. Once you start introducing that
56:21
into what I was talking about, it gets like 10 times more
56:23
complicated so that's it. Definitely batches what I had in
56:27
mind. So what does it look like?
56:31
So, I'm doing this on the fly. But what do you want to see what
56:36
a semantic layer is? See what you got.
56:39
Okay, I didn't prepare for this at all everyone.
56:42
So it's alright, this is bonus here.
56:44
This is like icing on the cake or this is the extra that you
56:48
wouldn't expect. So if people are even still here
56:51
with us, the we've got them and George.
56:57
Have you ever messed around? Have you thought about this or
56:59
you've heard of this? Yeah.
57:00
So I'm like a heavy user super sad and I completely believe
57:04
their model of leave as much of that as you can to DVT.
57:08
And even so I use Daxter. So if I need to define a data
57:13
set that I could only Define in Python, I would Define that with
57:15
software-defined Assets in dark. So, which I think overlaps with
57:18
a semantic layer. I just view the semantic layer
57:21
is being sort of management of what you'd normally end up,
57:26
displaying in a visualization tool or use an input to an ml
57:28
model. You want people to be defining
57:31
that in data as opposed to in endpoint notebooks where you
57:36
might start like having different definitions of The
57:38
number of users you want to like try to move that up in your
57:42
pipelines so that everybody can agree on what a date Point
57:44
actually is, and come to the same number when they need to
57:47
access it. And how do you see this playing
57:50
around with machine learning? Well, I mean, you can try too
57:56
often, you'd be able to define the, the input features to your
58:00
machine learning training and kind of conventional mortgage
58:03
consumer side. Machine learning say, like,
58:07
trying to predict the number of future use Is or things like
58:09
that, you probably do a lot of your data is data cleaning and
58:13
trying to construct, those metrics in raw, SQL management,
58:16
something like DDT and then you just have your and our pipeline
58:20
damage Downstream from that. And then, the advantage that you
58:23
get, is that you could consolidate that between your,
58:26
your ml pipelines and your visualization.
58:28
So, if you're trying to predict the number of users, you can use
58:31
that same number of users to find.
58:33
Once in a DVT, query user in your email pipelines for some
58:37
time series prediction. Ins and use that in your
58:40
visualization tool with something like super set where
58:43
you can clearly Define and make sure everybody agrees on what
58:45
something is and alongside that your thing, get sort of your
58:48
data catalog, which is, if you want to find some electrical
58:52
find some column, you can search through all of your data sets.
58:55
And as I sort of new data scientist, joining a team will
58:58
be your One-Stop shop for any date that you need amazing.
59:02
Alright. Byron.
59:03
What do you got for us? Okay so what you see here is a
59:07
demo we put together in just satellite in front of images and
59:11
we're basically turning it into a model that we can ingest
59:17
through looker and we're I'll kind of won't go into the use
59:22
case and that kind of stuff because it's not really
59:24
relevant. But you can see here, there's
59:29
lat-long to to do. There's like the centroid points
59:32
and you can see them scattered throughout this map.
59:35
And you can also see a number of different average averages for
59:40
in DWI value, which is a type of value that can be derived from
59:44
satellite images. So the thing here that I wanted
59:48
to show you was let's take for example, this Geo, hash level 8
59:54
centroid, right? Let's go see.
59:56
That looks like in the semantic layer.
59:58
So this is the semantic layer. It's just a bunch of rules for
1:00:02
the your varying Dimensions that you would present.
1:00:06
So, effectively, this is serving as how you would develop for
1:00:10
these different features. Now, you can also do a lot of
1:00:13
different certain things. So for example you can embed a
1:00:16
URL to into a feature. So when it gets surfaced in a
1:00:19
map, you can click on it and you can go to this URL.
1:00:22
That's been defined here. This obviously is not as ql, but
1:00:29
you want this logic? This business logic to be able
1:00:33
to push down to your database. So, what's done behind the
1:00:37
scenes? Is this is transformed into SQL
1:00:40
and so, it's already. It's already executed this
1:00:43
query. We can execute again.
1:00:45
You can see it just pulled from cash because I had already
1:00:47
executed a few minutes ago, and if you want to see what that SQL
1:00:51
queries. So here's the output.
1:00:53
Here's the visualization. And here is the s, Underlying
1:00:55
ask you out so that logic that I showed in this other tab here.
1:01:00
Generates this query that's what a semantic layer is endure in
1:01:06
this interpretation of it. You could, you could view a
1:01:09
different ways and what's the strong motivation to move away
1:01:13
from SQL in this sort of use case?
1:01:15
I just I started off from the kind of sequel side.
1:01:20
I find it much easier to think through and there's you know
1:01:23
endless Forums on how to write every single query you could
1:01:25
ever want. You can normally like Stock,
1:01:28
Exchange stack Overflow, it already is the NLP model to spit
1:01:31
out the SQL query for No problem.
1:01:34
It's over. That's a yeah, that's it. 100%.
1:01:37
A very fair observation. I started from SQL as well and
1:01:40
to be honest, when I was getting used to this, I had to go to the
1:01:43
SQL queries, underneath it to kind of better feel comfortable
1:01:49
with it because I was starting from a point of feeling
1:01:51
comfortable desk you up. But why I moved to this one
1:01:55
point is Version Control, so you can Version Control this.
1:01:59
You can also Version Control, SQL, Vietnamese, Yeah, that's
1:02:06
okay. So the logic that you see here
1:02:08
on the screen, it can be bundled together with other aspects that
1:02:11
have nothing to do with SQL. You also have logic that can be
1:02:15
embedded in here that again has nothing to do with SQL.
1:02:18
I cannot use an SQL model to embed this URL into into this
1:02:26
Dimension. However, in this singular data
1:02:31
semantic layer, I can so The business logic becomes close
1:02:34
together, so that would be, those are two reasons, one
1:02:38
Version Control, which is probably not the primary one but
1:02:42
easily it's obvious. The other is your logic can be
1:02:46
grouped together more efficiently when you're talking
1:02:48
about things that are SQL and other than SQL and the other
1:02:55
side of it is you can start to enforce some logic to protect
1:03:00
people from making silly mistakes, you can start.
1:03:03
Embed data test unit, test into the business logic itself rather
1:03:09
than running it in separate components.
1:03:11
So you're trying to bring as much into that semantic layer,
1:03:14
which is why I kind of my thinking is I think there's more
1:03:18
of a future for a thick semantic layer in the coming years.
1:03:24
So sorry we went when I was excitedly interjecting, it was
1:03:27
just because DVT it really does make managing sequel in
1:03:31
something like it, so easy. So my way of combining sequel
1:03:35
with other logic is that there's there's now Integrations between
1:03:39
DVT and most pipelining tools. So again this is all like
1:03:42
relevant to the batch processing side of things, much more than
1:03:45
any other context but there will be.
1:03:47
There's definite are you heavily used the dikes a DVT
1:03:49
integration? I suspect as the same with
1:03:51
prefect and airflow. And they so that you can
1:03:54
reference the output of a DVT SQL query in your sort of Dexter
1:04:00
python code very, very easily. So you can switch from python to
1:04:04
SQL as many times as you want to in a single Pipeline and then
1:04:08
you can manage the logic of what needs to run Upstream to get
1:04:11
your desired output. But I'm I think I'm earlier in
1:04:16
this process, you've gone from using SQL to then discovering
1:04:19
that semantically offers quite a lot.
1:04:22
So I need to explore. Yeah, well I mean I think what
1:04:26
you said is really valid like DVT is a powerful tool and it's
1:04:29
awesome. I do think it has kind of it to
1:04:33
me. It's value is more in the data
1:04:36
pipeline. So if we talk about, if we think
1:04:38
about data, any data work that builds up to any kind of and
1:04:42
data enrichment, there's three parts.
1:04:45
If you want to break it into its broad constituent components you
1:04:48
have in just you've transformed and then you have your final
1:04:52
layer where you're kind of pushing back out to some other
1:04:54
application. I think the semantic layer is
1:04:57
that I'm talking about has a really a home at that kind of
1:05:00
when you're pushing back out to some other application, whether
1:05:03
that's ml or something else. Whereas in the case of what like
1:05:09
DBT, I think it's home in my eyes.
1:05:12
Kind of really sits in that transformed that middle bit
1:05:15
where, where it is needed. And, you know, I'm sure there's
1:05:20
there's that the line of where One picks up in one.
1:05:25
The other one drops off is I think always kind of fuzzy and
1:05:29
changes any ways depending on these cases.
1:05:32
Excellent. This is super cool.
1:05:35
So, so that was my like, that's why I was really excited about
1:05:39
what maximum talked about because it kind of related back
1:05:41
to some some of my recent experiences.
1:05:45
Perfect perfect. Well, this is great guys, that
1:05:49
was the little bonus section and now we're really gonna to finish
1:05:54
it. That's all we got for today.