Latest / Agentic Conversations (formally mlops.community) / MLOps + BI? // Maxime Beauchemin // MLOps Coffee Sessions #104

Transcript

MLOps + BI? // Maxime Beauchemin // MLOps Coffee Sessions #104 · Agentic Conversations (formally mlops.community)

• 0:00
  Bishnu. You see all the kids toys behind
• 0:02
  me. Oh, yeah, where you at on
• 0:06
  officially in Toronto, for the ml ops world event?
• 0:10
  It is going to be in person. You are threatening to come.
• 0:15
  I'm too threatening. It's a firm threatened and you
• 0:17
  gotta send me the times. Yeah, we're going to try and
• 0:19
  record some live stuff at the event.
• 0:22
  So, should be cool. But today, we're talking with
• 0:26
  Maxine. Buscemi, Buscemi, you know how I
• 0:29
  am with the names and I cannot get them right for the life of
• 0:32
  me. But this guy You have not heard
• 0:35
  of them, you probably have used a tool that he has created,
• 0:39
  whether that was air flow or preset.
• 0:42
  The dude is all over the place. And sorry.
• 0:46
  Did I say preset? Is it preset?
• 0:48
  You since the managed company that he runs?
• 0:50
  But superset is the original tool, which is now run by the
• 0:53
  Apache Foundation. Yeah, there we go.
• 0:56
  That's what it was. I knew there was some and so I
• 0:58
  loved some of this hot takes around like why he would go and
• 1:02
  start a company around. Round.
• 1:05
  Like preset well, no. What is it?
• 1:10
  God, I can't remember the name of it.
• 1:12
  Now I just started getting started super set, which is
• 1:14
  opens, open source and then started preset.
• 1:18
  And the hot takes were about. Here we go, get a business
• 1:19
  intelligence. There we go.
• 1:21
  That's what it was. So I have a total mind blank
• 1:24
  right now. And thank you for walking me
• 1:26
  through that the other part, which we just drove right into.
• 1:30
  And I really enjoyed that we got right into to it was this
• 1:34
  entity. Eccentric type of thing.
• 1:37
  Maybe you, what did you take away from that?
• 1:39
  Like, because I know that you were loving it.
• 1:43
  Yeah, for sure. I mean my big takeaway and I'll
• 1:46
  go through Max Max's like by on a second.
• 1:48
  Everybody should should get to know who this who this man is
• 1:51
  and what he's done for all of us.
• 1:53
  But what we talked about was entity Centric data modeling.
• 1:57
  And what why that's important is increasingly organizations are
• 2:02
  are generating tons of data that we want stored.
• 2:05
  Putting that into the data warehouse and using it to drive
• 2:07
  analytics and machine learning outcomes.
• 2:10
  But how we make that repeatable, reliable, scalable and
• 2:14
  maintainable is through data modeling which not enough
• 2:17
  companies do well, that was Max's point and he gave us sort
• 2:21
  of a understanding of how to think about doing data modeling,
• 2:25
  in a modern context that isn't just hey, go read Kimble's books
• 2:29
  about dimensional modeling or like that stuff was written
• 2:33
  before we had like the MLS Ops and machine learning stuff we
• 2:36
  have today and he gave us an intro into like how to think
• 2:39
  about it in 2022 is feature stores and all the other sort of
• 2:42
  systems that are around now. This is cool.
• 2:45
  Yeah, let's get into it man. That was really good.
• 2:48
  Let me take you through. Oh yeah, 95i real quick.
• 2:52
  Yeah, for sure Max, you should follow Max pusherman because
• 2:56
  he's currently the founder and CEO of preset but more relevant
• 3:00
  to this podcast. He's actually the original
• 3:02
  creator of both Apache superset, a business.
• 3:05
  Halogens open source tool and Apache airflow, which is a data
• 3:09
  orchestration probably. All of you have interacted with
• 3:12
  and a he has had an exciting career shaped at places like
• 3:15
  Facebook and 2012 Airbnb in 2014 and lift and so many cool blog
• 3:20
  post, check it out and do not forget, we started a pretty
• 3:24
  awesome newsletter that you will want to also subscribe to if you
• 3:29
  have not as the best of slack or just join us in slack and get
• 3:34
  all the information. Formation drink from the fire
• 3:36
  hose, let's talk with Max. So max, you just have something
• 3:43
  pretty interesting and you need to talk about this idea of
• 3:46
  entities entities eccentric, data modeling and why that's an
• 3:50
  interest to you. And let me start by saying like
• 3:54
  I feel like a lot of machine learning is starting to meet the
• 3:58
  modern data stack and like people are trying to do machine
• 4:04
  learning in the day. The warehouse and are trying to
• 4:08
  apply Concepts like that before we get to like actually talking
• 4:12
  about data modeling. I wanted to get your thoughts on
• 4:15
  that Paradigm of just throwing a shit ton of data into the data
• 4:21
  warehouse and then doing machine learning on top of it right
• 4:23
  there. Does that make sense?
• 4:24
  Or are we just like setting ourselves up for a future
• 4:26
  problem? Yeah, I don't know.
• 4:28
  I think one thing that makes sense, it's the data as kind of
• 4:31
  gravity to do it, right? It's going to have the very
• 4:34
  heavy. Thing.
• 4:35
  And they don't want us to be together in one place and
• 4:38
  ultimately all data wants to be joined and uni, and and kind of
• 4:44
  brought together. So I think the bid to bring it
• 4:46
  all in one place and call it data warehouse, her legs or data
• 4:50
  Shack and you know whatever you want to call it make sense to
• 4:54
  take all your data and bring it one place.
• 4:56
  Where you can make sense of it then, does it make sense to use
• 5:00
  to compute resources inside the data warehouse to do amount?
• 5:05
  It's like bring the computer where the data is, bring the
• 5:07
  data where the computer is. I don't know.
• 5:10
  It seems like it does make sense if it is, you know, on a cluster
• 5:13
  on still flake or cluster and on bigquery cluster to bring the
• 5:18
  compute right there. It's easier to move to compute
• 5:20
  than moving the data then, you know, a sequel, the right set of
• 5:23
  semantics to express and now transformation or you started
• 5:27
  having this stuff where you have.
• 5:28
  Like I remember doing some that stuff at Facebook, where we add
• 5:31
  them, you'd ETFs and udf's and So you guys like and I already
• 5:36
  did user-defined, you know, functions inside your Sequel.
• 5:40
  And I think it's not that crazy to think you would have a
• 5:44
  program running in line in between like, SQL resources,
• 5:48
  very like, kind of Select star from you DTF.
• 5:51
  And that you DTF going to be, you know, whatever language I
• 5:56
  think. Like I was talking withdrew from
• 5:58
  DBT recently and he was saying like, oh I want like they're
• 6:01
  really interested in people running python code that you
• 6:03
  don't want to be running. You know, I think about in DBT
• 6:06
  itself, they want for other things to run the python for
• 6:09
  them. It's like, I don't want to
• 6:10
  python interpreter running random stuff in my, in my own
• 6:15
  kind of Stack, but I think we're getting there.
• 6:17
  Snowflakes, got some stuff I think, where you can run kind of
• 6:21
  python in line and things like that.
• 6:23
  So it doesn't seem that crazy to me, you know, who's been
• 6:25
  crazier, things done in the past.
• 6:29
  Yeah, you're definitely right that there's been crazier,
• 6:30
  things done in the past especially when we think about
• 6:33
  all the hacks that people have applied to data over the years,
• 6:37
  over the generations. And I think where that question
• 6:40
  came from. For me, as you know, I've been
• 6:42
  reading a lot of Chad Sanderson stuff.
• 6:44
  He's the director of data platform at Convoy and he writes
• 6:47
  a lot about how data modeling is more crucial than ever.
• 6:53
  We had him on gonna meet up recently, check it out to all
• 6:55
  our listeners, but he talks about how data modeling is more
• 6:58
  crucial. The never with the proliferation
• 7:00
  of data sets and the fact that they're all ending up in one
• 7:04
  place where sometimes too little structures applied.
• 7:06
  So that's where that came from. But I guess my question to you
• 7:10
  now is you mentioned entities Centric, data model data
• 7:14
  modeling, what does that actually mean?
• 7:16
  And how does that apply to ml this?
• 7:18
  It's an idea that I've been pushing forward, not talking a
• 7:21
  whole but bunch about or not even writing a blog post on
• 7:26
  basis. I was looking at ml
• 7:27
  practitioners and what they do and how they work and just like
• 7:31
  another training and what it requires.
• 7:33
  And really often, I think what you need is a really clear
• 7:36
  entity. And then a shit-ton of, you
• 7:39
  know, features and attributes and things.
• 7:42
  And some of these things are not just like an attribute of the
• 7:46
  data. So if you're, you know, doing
• 7:48
  some Trend predictions around a user, you might have a lot of
• 7:50
  demographics on that user. And that's interesting.
• 7:53
  But you also have a lot of facts and metrics and you know 7 David
• 7:58
  It's and 28-day photo upload on something like Facebook, right?
• 8:04
  And the number of like 364 day, bookings at Airbnb.
• 8:09
  And you want to take all of these features and then
• 8:12
  associate them with an entity and it ethic like, like stepping
• 8:16
  back and talk about data modeling.
• 8:17
  So it's a little bit of a lost art, right?
• 8:19
  Let's say we have a weird democratizing analytics process,
• 8:22
  and everyone is invited and everyone knows Sequel and, you
• 8:25
  know, airflow and DBT and but like no.
• 8:28
  No one really knows that lost our data modeling.
• 8:33
  Very, very well. So that's kind of an interesting
• 8:36
  thought and then getting into dimensional modeling that it's
• 8:40
  just for people. There's some books, you know
• 8:42
  written by Ralph Kimball from probably like you know, 30, 2530
• 8:47
  years, old books around like methodology on how to organize
• 8:51
  and structure your data, how to create the right data set?
• 8:55
  So that it's best organized for retrieval.
• 8:58
  Consumption rights. And dimensional modeling is one
• 9:00
  of these things where you have your typically you know you have
• 9:03
  what you call this star schemas or snowflake schema is and
• 9:06
  sounds very fancy. But but generally, you can think
• 9:09
  of these schemas adding facts, you know, like something like
• 9:13
  bookings at Airbnb. And in Dimensions, that are
• 9:16
  collections of entities things, like, you know, hosts and, you
• 9:22
  know, you have like, who is doing the booking?
• 9:25
  Whoo. What is the list thing for the
• 9:26
  bookings? For those Earth?
• 9:28
  Mentions and those dimensions are entity Centric in some ways.
• 9:32
  Are there that dimensions a dimension like user or listing
• 9:37
  is in a lot of ways, a an entity then dimensional modeling the
• 9:43
  available. What is a fact the city of fact
• 9:45
  tables and dimension tables. Well, we see in ml is like you
• 9:49
  know people want to have these that more feature for store.
• 9:52
  So you take an entity like a user can input all of the
• 9:54
  attributes including the facts as attributes.
• 9:59
  Of the entity and I really often very common practice is having
• 10:04
  like these time Windows. Like how many people?
• 10:07
  How many times did someone do this over a certain period of
• 10:10
  time? Which is usually down to now,
• 10:13
  right? Because I'll many many times you
• 10:16
  booked on Airbnb, last year's a great predictor.
• 10:19
  Perhaps of like, how many times you will book next year?
• 10:21
  So this very useful and, you know, machine learning models,
• 10:25
  may know how to make sense of these metrics and weigh them.
• 10:29
  The point being like, okay, so that's a thing now, call it
• 10:33
  feature store, but then I think beyond the idea of a feature
• 10:36
  store and Beyond the idea of making prediction and feeding ml
• 10:40
  model. There's a bunch of other reason
• 10:42
  why it's really cool to have and City Centric metrics, right?
• 10:46
  Like the future store is greater and it's use than it is to the
• 10:54
  to just like train models, right?
• 10:56
  So if you can have like a million attributes about a you,
• 10:59
  They're at Facebook precomputed when it comes time still
• 11:04
  retrieving. Say, I don't know scorecard for
• 11:06
  a user to be like a what's happening with this user?
• 11:08
  So that's it, there's there's there's a lot of value and
• 11:11
  knowing these things and there's a there's a lot of value in
• 11:15
  having the data organized and that way.
• 11:17
  And that's what I call. You don't have to T Centric
• 11:18
  modeling, where you have free, really clear entity, and then,
• 11:22
  you know, an Infinity of not only attributes that user about
• 11:26
  features in the sense that time window.
• 11:28
  Action and Mike, you know, metrics at that level 2.
• 11:33
  So it's a little funny because I remember visiting someone's
• 11:39
  office when I was in London and they were talking about how they
• 11:43
  have so much metadata now that they need metadata for their
• 11:47
  metadata. And it kind of feels like you're
• 11:51
  touching on that a little bit. Like he amount of stuff that is
• 11:55
  going on but this entity Centric.
• 11:59
  View will be a way around that. It's like you have the main
• 12:04
  protagonist and then everything else that goes on there.
• 12:09
  Could possibly be a remedy for that.
• 12:14
  Yes, it really interesting thought and, you know, the I
• 12:17
  think like that, the idea that metadata is growing is
• 12:21
  definitely super clear to me to write like this idea to of that,
• 12:26
  we need a meta data warehouse the same way.
• 12:27
  We have a data warehouse, we need to take all the metadata
• 12:30
  from the company, like information about, then image
• 12:34
  and data provenance and data Ops information, like, operational
• 12:38
  metadata. Like, you know, what was
• 12:40
  computed and one and by which grip, and who's the owner?
• 12:44
  Of that stuff. And where is it coming from?
• 12:47
  So business metadata to of like labels and definitions, we still
• 12:53
  only like have more and more of that, nothing like the stuff
• 12:57
  like Munson and data Hub. If you're familiar with these
• 13:00
  like open source project is trying to treat all that
• 13:03
  metadata like bring it on a central place when you feel kind
• 13:07
  of a graph of all your metadata and your organization or all
• 13:12
  your data assets and bring it all together.
• 13:13
  Together so you can see kind of neighborhood.
• 13:17
  So you look at a table, you know, where it's coming from
• 13:19
  what's upstream and downstream of it was the owner that kind of
• 13:22
  stuff that's of is like the more I think the more your data
• 13:27
  platform gets complex and ages and has history, the more
• 13:31
  important that is to like make sense to stuff the ways when
• 13:34
  you're stuck with like you know, tribal knowledge and asking your
• 13:37
  peers, like can I trust this table?
• 13:39
  You know, is that reliable? Yeah.
• 13:41
  Who owns this stuff? So that's, that's an interesting
• 13:46
  thought, you know, that metadata as data experiment and talks
• 13:51
  about like how do you know they exhibit B?
• 13:53
  You know, they were, there's like tons of data everywhere and
• 13:56
  it's like growing, it's potentially, I think it's, you
• 13:59
  know, metadata is similar. It was a challenge to.
• 14:02
  Yeah, it's your point about the data platform and its age and
• 14:07
  it's the relative knowledge and then, you know how it kind of
• 14:11
  becomes like hey you know, what's going?
• 14:13
  Hear what you think about this table?
• 14:14
  What is this feature? How does that actually computed?
• 14:16
  How are we actually Computing? Like, whether someone's
• 14:18
  defining, someone with this moment active or not, right?
• 14:20
  That's where all this like, tooling for tooling starts to
• 14:22
  evolve and, you know, the question I have for you, which
• 14:27
  is actually very much like my job.
• 14:29
  So I'm gonna ask you to do my job.
• 14:30
  Here is, let's say you're in a rapidly growing business and
• 14:36
  said rapidly growing business has mountains and mountains of
• 14:38
  data and very, very pressing business questions to answer Oh
• 14:43
  hey, we have a business model to figure out here.
• 14:45
  And you need to tell me whether or not my tab is what I think it
• 14:48
  is today. So you can answer.
• 14:50
  All these analytical questions doesn't feel like you have a ton
• 14:52
  of time to model the data and it doesn't feel like you have a ton
• 14:54
  of time to think about. Well what are the entities that
• 14:57
  map are rapidly to a map to a rapidly changing business model
• 15:01
  and say, said data, professional was in a team of two people with
• 15:05
  the team of two people. How would you go about wrapping
• 15:09
  your arms around this problem like data Organization for a
• 15:12
  rapidly scaling. All right.
• 15:15
  What's the first as I tell them is mostly, you know, an external
• 15:20
  problems that beset? Maybe that's not the best
• 15:22
  example, the Tammy just like, sure.
• 15:23
  Sure. Here I need specialized in like
• 15:25
  that us already done the analysis for the time of your
• 15:28
  Market. Well, let's say, what if it was
• 15:30
  a new market, that no one really knows about maybe an Emerging
• 15:33
  Market. Yeah, I think that becomes it's
• 15:36
  less of an internal and of the warehouse type question and
• 15:39
  maybe more of a research problem for the case of town, but let's
• 15:43
  Take. If if the question is more like
• 15:45
  York actu, LTV ratio customer acquisition, acquisition last /,
• 15:50
  lifetime value of customers. And pretty intricate complex
• 15:54
  like entry looking question inside your organization and you
• 15:59
  have to say, what is the vat, light, lifetime, value of a
• 16:01
  customer? You need to figure out, you
• 16:03
  know, look at your history and your customers and do some
• 16:06
  projections and Cactus much. Does it cost me to get a
• 16:09
  customer as pretty introspective of a question too.
• 16:13
  How do you do this? I think for me, I think
• 16:17
  historically or ones at the pattern that I've seen in the
• 16:20
  past is if you're in a rich, Jewish organization is you're
• 16:23
  going to do whatever you need to do to get an answer that
• 16:26
  question this week and that will typically will be, you know, a
• 16:29
  data analyst or I guess nowadays, maybe it's more an
• 16:32
  analyst engineer who's like, going to go and hit some raw
• 16:35
  tables and write a shit on a sequel and then maybe get a, you
• 16:40
  know, half these that answer to that question.
• 16:44
  And in fact, the last week. Yeah.
• 16:47
  Yeah, I didn't. Yeah, so that might be like a
• 16:49
  little bit better, put a finger in the air and then the best
• 16:52
  answer, I could get you this week, is this?
• 16:54
  And then and then hopefully this person, you know, and maybe is
• 16:58
  the same person, maybe it's a different person.
• 16:59
  But someone else is working on no closely with this person and
• 17:04
  they would be thinking about like, how do we do this, right?
• 17:06
  So we can get an answer to this question every every week for
• 17:11
  the foreseeable future and we can actually measure the And
• 17:14
  that's a little bit more of the data engineer Persona, who that
• 17:18
  tends to work at a much slower Pace, but does things right?
• 17:21
  Like builds things that are there, that will be there in a
• 17:25
  year or two from now and maybe five years from now, you know,
• 17:28
  we will still this person treats, you know, their art,
• 17:33
  like something their craft, or what they do, as a little bit
• 17:36
  more infrastructure that will have to be maintained in the
• 17:39
  future. And it may be is it the same
• 17:41
  person or not? Probably not, right?
• 17:44
  So one is probably vertically aligned and like, working on a
• 17:48
  weekly timeframe and the other one thinks more in terms of like
• 17:51
  maybe quarters and you know, and then use the source control and
• 17:56
  that's code review and thinks about data modeling and costs,
• 18:00
  and make sure that's everything that they've built so far, will
• 18:03
  lend every morning before 9:00 a.m. are, you know, has some
• 18:06
  essays and that sort of thing. So, the answer is like, at the
• 18:09
  end data, as in many, many things in fast, moving business
• 18:12
  has yet to do both. The short term and the long term
• 18:15
  solution in parallel. At the same time, the reality is
• 18:18
  like, do you always add have the resources to do both the often
• 18:22
  dry? So it's likely to give an answer
• 18:24
  this week and you put in your pilot project and maybe you
• 18:26
  don't get to it until something's on fire or you
• 18:30
  really do, you know, Seth or an executive's.
• 18:32
  Like I need my cocktail TV every every you know, every Monday
• 18:35
  morning, I'm kind of processing that answer because there's a
• 18:39
  lot there's a lot to dig into their and but you know what
• 18:42
  sticks with me? Is this idea?
• 18:43
  idea that like, yeah, you're right, somebody's got to be
• 18:45
  working on doing on fighting fires and somebody's got to be
• 18:48
  working on, sort of building the You know, building the
• 18:52
  foundation. So to speak, write and splitting
• 18:55
  that difference and putting two different roles there and having
• 18:58
  two different mandates to saying like we're going to be doing
• 19:00
  both in parallel is really important for I think data teams
• 19:03
  to both you know build and maintain trust and also like
• 19:08
  support the business for the long term, right?
• 19:10
  Because it's the twin mandate that you kind of have when your
• 19:13
  data professional. One thing I wanted to ask you
• 19:16
  about coming off of that is, you know, this conversation has kind
• 19:20
  of started off on a very In general sort of a data team data
• 19:26
  team focused, right? Like data Engineers analysts
• 19:28
  like what's what's generally happening in the modern data
• 19:31
  world and you know it sometimes feels like the machine learning
• 19:35
  world is like sometimes a subset of that and other times like
• 19:38
  whole world of its don't, you know?
• 19:40
  And you know they're kind of like slightly different forums.
• 19:44
  There's slightly different experts and I'm kind of
• 19:46
  wondering what your reaction is to that.
• 19:47
  Like do you feel like the way that we should think about
• 19:50
  machine learning tooling and stuff?
• 19:51
  It's just A subset or is it a little bit of its own?
• 19:55
  Yeah, it's there's, there's an interesting parallel to be drawn
• 20:00
  here. And if you saw that, you could
• 20:03
  say the future of data centers in our wits, came up with the
• 20:05
  modern data stack diagram, so that blog post and there, there
• 20:09
  was going to version 1.0 of it, which brought ML and did it like
• 20:13
  the whole data platform with MLS part of it in a single diagram.
• 20:19
  And then, for their 2.0 version of it, they can Did they decided
• 20:23
  okay, this is too complicated, it looks like at the spaghetti
• 20:25
  plate, you know? And then let's take the ml stuff
• 20:30
  and make it in its own diagram which is kind of interesting
• 20:35
  that I'm not sure if the fact that they did that, you know, as
• 20:37
  informative or you know, it'd be interesting to look at the
• 20:41
  differences between the two side-by-side.
• 20:43
  The, you know, I don't know if we could pull it here and and
• 20:45
  look at like what's the Venn diagram of that?
• 20:48
  Like what is on boats on both diagrams?
• 20:50
  And what is very unique to one? Or the other to get those relics
• 20:54
  of the data warehouse question, you asked before I like, so are
• 20:57
  we kind of converging or diverging?
• 20:59
  Here it is. Like clearly.
• 21:03
  It is different people with different concerns and different
• 21:07
  backgrounds operating on different time frames.
• 21:10
  I think on the ml side like notebooks are just really great
• 21:14
  which is what is a notebook, really.
• 21:16
  It's a program that you execute, you know and not Is the
• 21:23
  sequential order, statistic the repple diet, so the rebel cells
• 21:27
  like really works, I think it's important.
• 21:29
  And then for NFL practitioners and in the open-ended you know
• 21:35
  power of just like an interpreter as opposed to you
• 21:38
  know something like SQL is not sufficient, that's for sure.
• 21:43
  So yeah I think you know unfortunately I think it's it's
• 21:45
  at the answer is like it's an intricately shared resources
• 21:50
  across the two and then Some areas were there, like it's
• 21:54
  intricately, you know, diverging in some ways and if you think
• 21:58
  the funkier stuff, if you do, take the fun to your stuff on,
• 22:01
  the MLS that on the MLS side of the house, or that is like,
• 22:05
  let's say, like, you know, notebooks notes are also useful
• 22:09
  for data Engineers to write, like so it, maybe they would
• 22:11
  spend more time in SQL in general, and less time in
• 22:14
  notebooks. Back notebooks are still like it
• 22:16
  pretty great paradigms. If anyone you're trying to do
• 22:18
  there might not be an ml use case before more.
• 22:22
  Could things and did engineering, you might want to
• 22:24
  use no choice to. So, let's, let's pull on that
• 22:30
  thread a little bit more because it feels to me like, like
• 22:33
  airflow is very much a data engineering thing, but then a
• 22:39
  ton of data scientists and machine learning Engineers, love
• 22:43
  it. It's like some people have given
• 22:45
  talks on the community meet up where it's like air flow.
• 22:50
  The unsung hero of the The data scientists even.
• 22:54
  And so I wonder about the. Did you have that in mind was
• 22:59
  like when when you're looking at creating such a tool like that
• 23:03
  or when you're going through it, or you solely focused on a data
• 23:08
  engineer and then it just so happens that the data scientists
• 23:12
  also found a lot of value on it or were you doing research?
• 23:16
  Looking at what the data scientists use cases.
• 23:18
  Where to? Yeah.
• 23:20
  So there's this fair amount of History.
• 23:22
  Be there to select first, the inspiration for air flow.
• 23:26
  So, no air flow as I started at Airbnb 2014.
• 23:29
  Actually, I started in between jobs and between my job like
• 23:33
  Facebook, and I taken the job, you know, Airbnb.
• 23:36
  I talked with the team and it was really clear that I was
• 23:41
  going to work on something like the airflow because they didn't
• 23:44
  really have anything like that. And then I was coming out of
• 23:47
  Facebook, where Facebook ad had this Cambrian, explosion of,
• 23:51
  like, Inventing the whole data platform from the from scratch
• 23:55
  because the data the scale of the data were just too big to
• 24:00
  fit and Oracle or teradata database, right?
• 24:03
  So then they had invented them things like higher than Presto
• 24:07
  or had been in the making while I was there.
• 24:10
  But I build their own visualization tools.
• 24:13
  On top of that. I built their own kind of data
• 24:16
  portal at the metadata repository.
• 24:18
  I was talking about we were we had built one of those, we had
• 24:21
  had problems. Really like I would imagine a 5
• 24:24
  to 10 different schedulers and you know are like that's
• 24:28
  processing come that you generate computation type things
• 24:32
  like air flow and then when there was there were 22 of those
• 24:37
  had been emerging internally at Facebook, one called this one
• 24:40
  arm and one called day to be. And they're both, you know, a
• 24:43
  way to run that's jobs like bags of bad of that dags of batch
• 24:49
  jobs, which were expressed as new.
• 24:52
  No Sequel and or python. So so at that point you know and
• 24:57
  it's warm Facebook had been used for a general comment like
• 25:00
  General competition was used by their scientist was used by all
• 25:04
  sorts of people wearing all sorts of different hats just to
• 25:07
  do like data pipelines, which included bunch of emelda prepuce
• 25:12
  cases and probably is improper, you know, training as well.
• 25:16
  So I think like I knew so what do you think about what airflow
• 25:20
  is is just Either Express dags of batch jobs, with that are
• 25:26
  kind of time bound and you know, airflow will just like allow you
• 25:31
  to express sets of jobs that depend on each other and provide
• 25:36
  guarantee about the order of those jobs and one are going to
• 25:39
  run. So that's generally when you
• 25:41
  think about, it's not even like a data pipeline, specific thing,
• 25:46
  where it's very generic. Like it's a scheduler, you know,
• 25:50
  and an orchestrator. Stir.
• 25:52
  And then what you're going to do in there, you know, as kind of
• 25:55
  free to like limited only by you know what you can trigger and
• 25:59
  python which is like you know, pretty much anything.
• 26:02
  So that I knew it was going to be used in all sorts of ways but
• 26:06
  you know we were talking earlier about all the funky stuff that
• 26:11
  people do you know in the history of data and I the
• 26:15
  airflow certainly enabled people to do all sorts of things and
• 26:19
  you know at whatever ways. Now there's a real That article,
• 26:22
  if you saw it, there's a article called that unbundling of
• 26:25
  airflow people. I found it.
• 26:29
  Well, yeah, it's actually pretty, it's a cool article
• 26:32
  because it shows like a all this stuff that you've been like,
• 26:36
  doing an air flow which is like generate computation generally
• 26:39
  around data, you know, all these things that you've been doing
• 26:44
  right now are getting like cater to and in a more directed way,
• 26:48
  right? So if you're doing a maybe you'd
• 26:50
  better do that and give you T if you're doing Data sync and bring
• 26:53
  data from her staff services. Then you should just, you know,
• 26:55
  five trainer are bite. It's got an issue, interesting,
• 26:59
  take on it against things specialized, there's like
• 27:03
  services that emerged at do certain things very, very well
• 27:06
  and an airflow is pushed back to become much more of an
• 27:08
  orchestrator, and less of a compute engine, like, generic
• 27:13
  batch compute engine. I'm Emanuele.
• 27:16
  Mason machine learning engineer at stripe and author of building
• 27:20
  machine learning powered What if you don't want your machine
• 27:24
  learning models to explode? Well, you should subscribe to
• 27:27
  this podcast. So I am, I have to ask before
• 27:32
  Vishnu jump in was and you've probably been asked this a
• 27:37
  million times. So forgive me for just following
• 27:40
  the crowd as as you see astronomer doing what they're
• 27:45
  doing like that was never a thought in your mind.
• 27:48
  Or did you analyze trying to do like a managed airflow and you
• 27:52
  decided not to how did that play out?
• 27:54
  Oh yeah. So, so for context to for people
• 27:57
  who don't necessarily know me and the history said.
• 28:00
  So I started airflow, I started another project called superset,
• 28:04
  so Apache airflow Apache. Super said I started both of
• 28:07
  those at Airbnb in 2014-15. My passion really wanted to work
• 28:12
  super sad because it's visual it's interactive.
• 28:15
  It's color full. I'm kind of you know I wanted
• 28:21
  this rub business. Halogens and I want to change
• 28:23
  the way that you know, Mortals work with the data every day and
• 28:28
  that's just like a much more or at the time and now start things
• 28:32
  more interesting to me than data pipelines and orchestration.
• 28:36
  So after like a maybe two years of airflow is like okay I build
• 28:40
  this thing. You know I'm going to let it fly
• 28:42
  and let it grow and my passion really went to Super set and you
• 28:45
  know ended up starting a about three years ago.
• 28:48
  I started a commercial open source company around soup.
• 28:52
  For said called preset. And that's really just going to
• 28:54
  following my passion and interest and the stuff I wanted
• 28:58
  to work on. And I also like the risk-reward
• 29:00
  I think is more interesting or its higher risk higher reward on
• 29:07
  the dative is you know data consumption layers.
• 29:10
  How about like 4:00? Let's yeah let's talk about it
• 29:14
  but first like four people don't know what superset is urge you
• 29:18
  to check it out, right? So superset is an open-source
• 29:21
  pretty Say data consumption, dashboarding data, exploration
• 29:25
  data visualization platform. It's very much a competitor to
• 29:29
  things like, you know, Tableau and liqueur so that all the
• 29:31
  things that you can do with Tableau in look here you can do
• 29:34
  with on the open source front with super said we've been
• 29:38
  working super hard on it so it has become really awesome and
• 29:42
  really competitive the risk-reward ratio is higher to
• 29:47
  me because it's such a very like red ocean.
• 29:52
  Market. And there's just so many vendors
• 29:54
  and business intelligence harder to get open source, has to come
• 29:58
  in and disrupt, because there's like, very, very established
• 30:03
  vendors where orchestration was kind of this little bit more of
• 30:06
  a new segment. So I did think, you know, about
• 30:09
  starting a, you know, to go the more the direction of airflow
• 30:14
  and you know, investors kind of came to me and I hate we don't
• 30:17
  you start an airflow company and my heart and passion was very
• 30:21
  much a gun a superset. And then I've been really
• 30:24
  friendly and following closely the folks at astronomer so they
• 30:28
  they tried to get me to join too early on.
• 30:33
  I was talking with them up super early on and you know I'm very
• 30:37
  friendly with them and love it invested to some and an advisor.
• 30:41
  What them too. So I still have passionate talk
• 30:44
  about this stuff and being close to that world.
• 30:47
  But my really focus has been on the data consumption.
• 30:50
  Layer did a Ation dashboarding and how people like make sense
• 30:56
  of data, as you know what I've been working on.
• 30:59
  So you mention it is a red ocean that you're getting into like in
• 31:06
  a very blunt wait like what makes you think that you can
• 31:12
  disrupt this field? Well, so death takes, there's
• 31:16
  like two big angles and I think like through Innovation, you
• 31:20
  know, I think we can disrupt and Startups that have been
• 31:23
  disruptive but I would say there's two things that that we
• 31:26
  do that are very disruptive one is open source that like open
• 31:29
  source is just like this tidal wave of disruption, right?
• 31:34
  It comes it comes in and what it's, you know, it's free as in
• 31:41
  Freedom and free as in beer to so that means that you can just
• 31:44
  pick it up. Use it and get value off of it.
• 31:47
  I'll disruptive is that, you know, that's like a version of
• 31:49
  Freedom. You can also like extended you
• 31:52
  You can you can you can even build you know, integration
• 31:55
  points and we can extend things, you can make it your own.
• 32:00
  That's extremely powerful. I think people are pretty fed up
• 32:04
  with vendor lock-in to, and proprietary languages and tools.
• 32:09
  And, you know, just being at the mercy of companies, now, like
• 32:14
  Google cloud and on a side of looker and sales force on the
• 32:18
  side of Tableau. So you don't I think You know,
• 32:22
  people have a new way to discover validate and pick and
• 32:28
  ramp up on software and that new way is, you know, closer to what
• 32:33
  open sources or SASS. And that's the other, the other
• 32:36
  angle of like, how we can disrupt in this market, believe
• 32:39
  it or not is just like some stuff that comes with SAS like
• 32:42
  freemium. So, you know, preset offers free
• 32:46
  version of super set up to five users, for instance, and you can
• 32:49
  just like prove value and use it ran.
• 32:52
  And up to five users and make sure that it works for you
• 32:56
  before you even talk to a sales person.
• 32:57
  You might not even have to talk to a sales person ever, which is
• 33:00
  fantastic. So I think like, you know, all
• 33:03
  the things that come from SAS to things like time to value in our
• 33:08
  case like time into first dashboards time to dashboard,
• 33:10
  he's connecting to things that are really focusing on that
• 33:14
  early Journey user experience. And then using their mechanisms,
• 33:19
  like re-engagement emails. And things like being able to
• 33:24
  send people emails saying, here's some dashboards.
• 33:27
  You might like in your workspace or in your, you know, that
• 33:30
  people in your team, I've created.
• 33:31
  So the things that the Macan isms of, you know, your typical
• 33:36
  SAS company as applied to data visualization VI make total
• 33:40
  sense to and we'll make it more more virata.
• 33:43
  More variety to know? Yeah, it's, you know, this
• 33:47
  conversation went exactly the direction.
• 33:49
  I was hoping despite my complete lack of involvement, so, Oh,
• 33:52
  it's outstanding to see that but, you know, I guess what I
• 33:57
  want to share here is I work at a company where we struggle a
• 34:01
  lot with how to do business intelligence, right?
• 34:03
  And what that means is not necessarily that we can't figure
• 34:07
  out whether to buy Tableau or power bi but really that we are
• 34:10
  like we want business intelligence to work in a
• 34:12
  fundamentally different way than it has been, which is like, put
• 34:17
  up a dashboard of kpis and check it every day.
• 34:19
  And over time, people stopped paying attention.
• 34:22
  Six weeks and people are like is that even around?
• 34:24
  And you know we've had that experience was like quick site
• 34:28
  and you know the other sort of standard stale offerings.
• 34:31
  You could say that are out there and I think for us what we find
• 34:34
  is that you know, there are really two use cases, right?
• 34:37
  One. It's just like just monitoring
• 34:39
  to make sure that certain business things that need to
• 34:42
  happen are happening. And then the other thing which
• 34:45
  is like more insight driven which is, you know, what are
• 34:49
  some things we should be doing or what are Something we're
• 34:52
  aware of our need to explore and get ahead of before it's
• 34:56
  happening, right? And I would really say like we
• 34:58
  want to empower that ladder piece a lot more but really
• 35:01
  struggle with it and I was wondering you know from your
• 35:03
  standpoint from your Vantage Point like you've been at like
• 35:06
  your Facebook Airbnb, your now, helping no commercial as
• 35:09
  superset and bring it to the world.
• 35:10
  Like what are some of the solutions that were the ways
• 35:14
  that companies should think about that ladder problem of
• 35:16
  empowering, that sort of exploratory work and getting
• 35:18
  Mortals superpowers with data? Yeah.
• 35:22
  You know, there's your you were talking about like one angle,
• 35:26
  which is like, how does data inform even then, you know, the
• 35:29
  intuitive part of like, what we should be working on or what we
• 35:33
  should not be working on Based on data.
• 35:35
  But the way, one way to think about it is that things like
• 35:38
  data is kind of a dimension of all the work that every worker
• 35:43
  does every day. So if you're starting your work
• 35:47
  out, a new product within your company, like a news, the new
• 35:50
  SKU or new feature. ER, if you're or essentially for
• 35:56
  any, whatever role inside a company, whatever workflow,
• 36:00
  whatever process that data is a dimension of that.
• 36:03
  And really often, we don't use data to power like much of the
• 36:07
  processes and much of the role. So you might work in customer
• 36:10
  success and at a company, right? And not be data-driven.
• 36:15
  I'm just pick up your tickets and get on the phone with people
• 36:18
  and help them. But there is a huge data
• 36:21
  dimension of This yes, work, right.
• 36:23
  And you could ever get really intricate set of dashboard
• 36:27
  around. Not only what?
• 36:29
  This? Yes.
• 36:30
  Team a customer success team does every day, but also like
• 36:34
  the new things they might want to try our that they might want
• 36:36
  to do. So there's like, you know, so
• 36:39
  essentially the data Powers every role and every processes
• 36:44
  and there's a question like how do we make sure that, you know,
• 36:48
  all of the data that is needed by most of the people?
• 36:52
  EBay and organization is available to to create the
• 36:56
  validation. And the opportunity is validate
• 37:00
  the Intuition or invalidate some intuition.
• 37:03
  Really often that did is just not there, right?
• 37:05
  We don't have the data is just like, it's engineering, like, we
• 37:07
  don't really have the right now. What to do.
• 37:11
  So yeah, I don't know if this is like insert answering that
• 37:15
  question that I was like how do we become more data-driven?
• 37:17
  I think like it's there's a hunch of like how do you show up
• 37:20
  to a meeting with like the State question of like a.
• 37:23
  I wonder what our percentage of you know our success rate for X
• 37:28
  is then it's like well, instead of just wondering about it went
• 37:32
  on someone, I can go and get that data and bring it to the
• 37:35
  next meeting. Or maybe it's the data is
• 37:38
  already there, so we can validate validate hypothesis
• 37:41
  instead of it as just like getting my intuition versus your
• 37:44
  intuition. Is this not super Concrete in
• 37:46
  useful? So these are these are
• 37:48
  organization that I was at are like lift Airbnb.
• 37:52
  Facebook Yahoo that. I think you'd have a lot of
• 37:54
  that, culture of like, I wonder, you know, about X.
• 37:58
  And let's get the answer to that question and live.
• 38:00
  Let's get that data and bring it to the table.
• 38:03
  It's not accepting, you know, raw intuition without data.
• 38:09
  Yeah, it makes sense and it's the cliche, right?
• 38:13
  Like the data-driven organization and you hope that
• 38:16
  most organizations are doing that.
• 38:19
  And so changing gears a little bit.
• 38:22
  It. You spoke about how you play in
• 38:26
  the data engineering field quite a lot, but you also are building
• 38:30
  for data scientists and machine learning Engineers because a lot
• 38:35
  of the stuff that you do inevitably touches on their
• 38:39
  space to are there things that you found over time as a builder
• 38:46
  and the data engineering space that you need to watch out for
• 38:50
  when it comes to building. For data scientists in terms of
• 38:55
  tooling, or did a pipelines or both?
• 38:58
  Yeah, I would. Let's start with data pipelines,
• 39:01
  but if there's something that's calling at you, for sure, jump
• 39:05
  on that. Well so I think like, you know,
• 39:09
  big part of the science is you know making models and big part
• 39:13
  of making models is training models and big part of like
• 39:16
  trading model is about like prepping the data to train your
• 39:19
  models and being rigorous food. To be a rigorous around that
• 39:24
  which you should be, especially given, you know, the impact that
• 39:28
  we see in that, you know, ml can have on and what in the world
• 39:32
  like you look at the Facebook like, you know, news fit
• 39:35
  Newsfeed ranking algorithm and the impact that has had on like
• 39:40
  parallelization and our society. You know, you're like okay the
• 39:44
  way that we train this model or this machine learning Beast
• 39:51
  seems Is really, really important and, you know,
• 39:54
  provenance and reproducibility for something like that.
• 39:59
  Seems really, really important. So that means like a lot of the
• 40:02
  best practices. I think to make sure that data
• 40:07
  engineering is done a right does apply to data prep for training
• 40:11
  models. So in concerns around reproduce
• 40:14
  the bowl of reproducibility Province.
• 40:18
  What was this mod? What was this machine learning?
• 40:22
  All trained with and can we reproduce that?
• 40:24
  And then becomes very important. So I would say the work that I
• 40:31
  did around sites, I did some, I gave some talks and wrote a blog
• 40:35
  post on something called functional data engineering,
• 40:37
  which is applying some of the functional programming Paradigm
• 40:42
  to date and real. It's like trying to draw some
• 40:44
  parallels between the two and seeing where these parallel
• 40:47
  makes sense, then there where their virtuous and And and then
• 40:53
  encouraging people to go in that direction to say like a or jobs
• 40:57
  should be pure functions, your tasks, you know, when you do
• 41:01
  pipelines to be liked or functions, you should use
• 41:05
  immutable data blocks and create other immutable data.
• 41:08
  Bluffs, and you should have like deterministic results.
• 41:11
  Like when you run this function, with this immutable data block,
• 41:14
  you should always get this result so that you can get to
• 41:17
  some level guarantees around reproducibility or at least be.
• 41:22
  Able to explain their results that you got.
• 41:25
  And it's like, that's important. So, the rigor and data,
• 41:28
  Engineering Process does apply to everything related to like
• 41:31
  preparing data, sets to train models.
• 41:34
  Yeah, I actually I love that point.
• 41:36
  The last point you made you for those of you who haven't read
• 41:39
  them, go check out Max's blog posts.
• 41:42
  Especially you know, that one about functional programming in
• 41:44
  the data context. Like there's so much of it work
• 41:47
  that we do as data professionals and is machine learning
• 41:49
  professionals is just applying the principles of software.
• 41:51
  Nearing to data, right? Like that's what it's about a
• 41:55
  lot of times it's like figuring out where it works and where it
• 41:57
  breaks right? And it's a really cool blog post
• 42:00
  that I would check suggest checking out and with that I
• 42:04
  want to get into some rapid fire.
• 42:05
  Some lightning round stuff. We got some we got some stuff
• 42:10
  lined up for you. I want to go first what was the
• 42:13
  craziest crazy man? Have the right programs you use
• 42:16
  it, craziest work environment for you to walk into Facebook
• 42:20
  and 2012 Airbnb in Or lift in 2017.
• 42:23
  So Define crazy. But I would say like the most
• 42:26
  surprising and changing was Facebook into in 2012, huge
• 42:31
  environment and reinvented everything from the ground up,
• 42:35
  super Dynamic, no rules. And by the same time like
• 42:40
  tremendous and like amount of amazing work taking place so
• 42:45
  that that was before we knew that Facebook or before a lot of
• 42:48
  the issues that came up with the facebooker to is like I was
• 42:51
  Before there was all that, this doubt in the air.
• 42:54
  But, um, you know, moving fast and breaking things and really
• 42:58
  living by. It was really impressive and,
• 43:01
  and power and power. See like I'd never seen anything
• 43:05
  like it, and never imagined that it work.
• 43:08
  Replaced could work with like, loose rules.
• 43:11
  Like that was pretty and amazed to be part of it and really
• 43:14
  empowered by it to drink Kool-Aid.
• 43:17
  For sure. I gotta say like those three
• 43:20
  things is on your on your resume.
• 43:22
  Times that it's at. It's kind of like playing for
• 43:23
  like Barcelona then Real Madrid. Then like you know, Manchester
• 43:28
  City, like one. Actually the next question that
• 43:31
  I have rapid fire is which of your blog posts.
• 43:35
  Were you most surprised by the reception to goodness?
• 43:39
  I think I like the rise of the day engineer was like super well
• 43:43
  receive, I'm not really surprised by it because I think
• 43:46
  I had to read maybe five years before the rise of the data
• 43:50
  scientist was like I'm gonna do this.
• 43:52
  For the data engineer, you know, and I think ideas, like likely
• 43:55
  to be fairly big because I knew that the role was going to be a
• 43:58
  big role. He starts to get her to data and
• 44:02
  interacting together as just such a powerful.
• 44:04
  You know, emerge of concept that I knew was going to be big but
• 44:09
  but it's still quoted many. What's surprising is just like
• 44:11
  the along and and how like a the life cycles it seems pretty only
• 44:18
  I would have thought like, well people read it.
• 44:20
  You get some comments in the first.
• 44:21
  Like few months and then it fades away, never talk.
• 44:24
  No, nobody talks about it again now.
• 44:26
  It's still like people still read this blog post.
• 44:29
  Yeah. Okay.
• 44:31
  So what is a piece of technology that would surprise people that
• 44:36
  you're bullish on? Well, that's a that's a tough
• 44:41
  one. I'm not sure if I feel like I
• 44:43
  need to really scan through my head, all the things that are
• 44:47
  controversial, you know, maybe I'll bring something, that's
• 44:50
  controversial. And bring up the angle that I
• 44:54
  like. But I don't know if you've heard
• 44:55
  about data mesh and it's been controversial on like data
• 44:59
  better and now aware because I think like, it's not very
• 45:04
  prescriptive in terms of like Solutions and what it means for
• 45:08
  data modeling or for today two products, you know what really
• 45:15
  it means. But I think it does a really
• 45:16
  good job at summarizing the issues that we have around the
• 45:21
  day. Governance and you know how
• 45:24
  people should work with data and modern organization.
• 45:28
  There's new to YouTube point, you know?
• 45:30
  It's like people don't really know what are the what are the
• 45:33
  roles than what are they approach is and how do we do
• 45:36
  data, right? The, you know and large are fast
• 45:38
  growing companies so I think they Identify some of the
• 45:41
  challenges very well. I don't think they offer like
• 45:43
  really clear solutions to it. But they do a good job at summer
• 45:49
  I think on the issues of our time.
• 45:50
  So you know, I do Data that a team culture, it's hilarious.
• 45:55
  You say that because we when we've had people on here that
• 45:58
  talk about data mesh, I liken it to World Peace.
• 46:03
  It's like yeah everybody's for World Peace like don't know,
• 46:06
  Buddy's not saying we don't want world.
• 46:08
  Peace is just like, how do we go about that and how do we get it
• 46:12
  like in actually execute on it? That's another story.
• 46:15
  Like it's kind of in that I throw it in that field of my
• 46:20
  brain. So All right, I got more on
• 46:23
  that. It's like when you read the
• 46:24
  paper you like, oh my God, this is so good.
• 46:27
  Like I relate to everything they say.
• 46:29
  And I can't wait to do to get to the chapter, where they offered
• 46:32
  the solution, and solve it all, and then he read it, it's like
• 46:35
  treating data like products and like, okay, that's kind of what
• 46:38
  I've been doing forever, right. And now and it's like all they
• 46:42
  just rename this concept of different name and that still
• 46:45
  does its kind of Falls flat a little bit on.
• 46:47
  Yeah verified. It calls out the problems and
• 46:51
  the pains Really well. And speaking we were talking
• 46:54
  about Chad Sanders earlier, and he's very much drink that
• 46:58
  Kool-Aid. And I think he's one of the ones
• 47:01
  in my eyes, that is exploring how to do it and what the
• 47:06
  remedies could look like at his own job he's trying to like push
• 47:10
  really move the needle and so that's that's a cool one.
• 47:15
  Alright. Last one for you and I'm torn
• 47:19
  between whether I want to ask you what the last book you read
• 47:22
  Or what the most ridiculous piece of marketing around data.
• 47:25
  You've seen has been, so I'll leave it open for you.
• 47:30
  Oh, goodness, that I mean, I think that the message might be
• 47:34
  like, it's a little bit, like every kind of a Consulting, like
• 47:37
  marketing play or, you know, there's a little bit of like,
• 47:41
  you know, naming things kind of land grab there too.
• 47:44
  So I think that we go there, I picked up a book that I just
• 47:49
  started, and now I've got to go look up the exact site.
• 47:51
  Tell, but it's about the the the life of Frank Zappa As Told by
• 47:56
  is like life partner and I do know that because I have to go
• 48:01
  and dig out the exact name. I read like the first chapter
• 48:04
  and not like take my life with Zappa.
• 48:07
  It's something like that. I'm just really interested in
• 48:09
  Zappa is like, you know, a musical genius kind of very
• 48:15
  interesting character to see the documentary.
• 48:18
  The recent one that came out. What was it called?
• 48:22
  Yeah, what was it called? Know like yeah, maybe but I'm
• 48:28
  gonna go look at my Amazon purchase list.
• 48:31
  There you go. But we can put in the show notes
• 48:34
  like let's just thrown on the documentary was great.
• 48:37
  There's something amazing about is that he's been kind of a
• 48:41
  hoarding as on material so you've been recording
• 48:46
  everything. So it's got like a is like he's
• 48:50
  got like a Like somewhere and is basement and one of his houses,
• 48:54
  it's got all the tapes and all the movies and all of the
• 48:58
  everything we ever recorded. So we can expect a lot more of
• 49:02
  the treasures to emerge out of that.
• 49:05
  Those are guys, which is amazing.
• 49:07
  I'm sure you've seen Zappa Plays Zappa Yeah, there's no result is
• 49:12
  a delete. Yeah, I haven't seen it.
• 49:14
  I know, I know it exists unless big andand Wizzle, you know,
• 49:17
  really? I'm sure it's a, yeah, I just
• 49:20
  don't know. I'm don't know him as much.
• 49:22
  It's crazy. How much so Dweezil, like looks
• 49:25
  like him and like acts like her and everything is so crazy.
• 49:29
  How similar it is. I like yeah dude his guitar
• 49:33
  playing skills. Are ya off the chart?
• 49:35
  Like like Zappa, like Frank. But I went to a few and they get
• 49:41
  the old legends that used to play with Frank to come on and
• 49:45
  just like tour with weasel or make some guest appearances.
• 49:50
  So I saw Return to Forever and and then Dweezil play with like
• 49:56
  Zappa Plays Zappa. Yeah, it's definitely worth it.
• 50:01
  Yeah, yeah. Yeah.
• 50:10
  Is. But American genius.
• 50:12
  Yeah. And check it out and maybe it is
• 50:14
  for you, if you haven't really tried, you know, what's a good
• 50:17
  Gateway album for people to start Joe's Garage, is my
• 50:20
  favorite Joe's Garage 80s, more like fun on the funny side.
• 50:26
  Yeah, by so many good album for the good jazz people are a
• 50:29
  little bit more coming from jerez.
• 50:31
  There's a Grand, Wazoo and wife it junghwa.
• 50:33
  Yeah, I really like hmm freak out and start with these
• 50:38
  recommendations. Well, And then it started.
• 50:41
  Stop a podcast. You don't speak out on that.
• 50:45
  Well, I mean, I could right now that's a so funny.
• 50:48
  You're saying this because one of my favorite things in that
• 50:52
  documentary that they talk about is how they asked him like, why
• 50:56
  are you doing this? You're spending all this money,
• 50:58
  you're going overseas. You're like Contracting
• 51:00
  orchestras, to make your music and he's just sits and they're
• 51:04
  like, and you're not going to make any of this money back.
• 51:07
  And he tells them straight up. He's like if it's music.
• 51:10
  That I want to listen to I've won.
• 51:14
  It's just like that's how I want to live my life.
• 51:16
  Like if it's worth doing because I will enjoy it, then I want to
• 51:21
  do it and I want to make it happen principal delivered by.
• 51:24
  Everybody says their new science.
• 51:26
  You want to consume or I don't know if there we go.
• 51:30
  We can extrapolate that into a moab's world.
• 51:33
  Maxine has been amazing. Man.
• 51:36
  Thank you so much for coming on here and sharing your wisdom and
• 51:39
  advice. To us all awesome.
• 51:41
  Once I give us pleasure to be on the podcast here.
• 51:44
  Thanks for coming.