WEBVTT 1 00:00:09.480 --> 00:00:13.119 Welcome to Alixer Mix, your Elixer podcast talking with members 2 00:00:13.160 --> 00:00:15.960 of the community. My name is Mark Erickson, and today 3 00:00:16.000 --> 00:00:19.160 we're joined by Sophie de Benedetto. Hey guys, and our 4 00:00:19.239 --> 00:00:20.800 guest Alvisa Sousmel. 5 00:00:21.320 --> 00:00:23.280 Hello, thanks for having me. 6 00:00:23.640 --> 00:00:24.640 I'm glad you could come on. 7 00:00:25.079 --> 00:00:28.000 So this is an exciting topic for me because it's 8 00:00:28.000 --> 00:00:31.120 something I find fascinating. I think whenever I hear about this, 9 00:00:31.199 --> 00:00:33.799 I think, Wow, that's super advanced and futuristic, and I 10 00:00:34.000 --> 00:00:36.159 have no idea how to do it myself. And so 11 00:00:36.240 --> 00:00:39.960 we've invited Alvisa on to help talk about how to 12 00:00:39.960 --> 00:00:43.119 do this, and so like what we're talking about is 13 00:00:43.679 --> 00:00:46.679 using like image recognition with the Elixer. 14 00:00:46.719 --> 00:00:47.960 And really cool stuff like that. 15 00:00:48.159 --> 00:00:52.079 So first of all, Alvisa, maybe you could tell us 16 00:00:52.119 --> 00:00:54.759 like where you work and a little bit about yourself, 17 00:00:54.920 --> 00:00:56.280 what kind of problems you're solving. 18 00:00:57.000 --> 00:01:01.159 Yes, sure, Seeing the last five six years I've been 19 00:01:01.320 --> 00:01:05.079 a CEO in a London hedge fund and I'm still 20 00:01:05.120 --> 00:01:09.319 working from them as a consultant and at the moment 21 00:01:09.359 --> 00:01:13.359 I'm consulting and running queticoding dot com, which is an 22 00:01:13.439 --> 00:01:17.920 educational site which is at the moment is mainly focused 23 00:01:17.959 --> 00:01:21.799 on Elixir and Phoenix, and I'm working to make some 24 00:01:21.959 --> 00:01:27.480 free and paid Elixir Phoenix courses. And my current clients 25 00:01:27.519 --> 00:01:33.239 are from finance, so the problems I'm facing are similar 26 00:01:33.319 --> 00:01:35.879 to what I've faced so far at the edge fund, 27 00:01:36.519 --> 00:01:40.640 and so to describe what kind of problems I had 28 00:01:40.680 --> 00:01:47.040 to face at the hedge Fund, We've designed three different architectures, 29 00:01:48.319 --> 00:01:51.480 mainly process data in real time. So the first was 30 00:01:51.560 --> 00:01:55.959 the process data for real time traits coming from different 31 00:01:56.000 --> 00:02:01.879 texchanges for hundreds of different products, products, our financial products, 32 00:02:02.040 --> 00:02:06.879 different assets, and the goal of this platform is to 33 00:02:06.959 --> 00:02:10.800 use the data to run our trading models. And in 34 00:02:10.800 --> 00:02:14.560 this case we had we have a lot of Python 35 00:02:14.840 --> 00:02:19.639 code because all the data science framework there are really 36 00:02:19.639 --> 00:02:20.479 great data. 37 00:02:20.680 --> 00:02:22.520 Science framework there. 38 00:02:22.919 --> 00:02:28.719 And another platform was is to process financial tweets in 39 00:02:28.759 --> 00:02:31.759 real time, so we get a huge amount of tweets 40 00:02:31.800 --> 00:02:34.719 in real time. We run our machine learning models to 41 00:02:34.759 --> 00:02:39.759 cluster these tweets to understand which company they refer to, 42 00:02:40.240 --> 00:02:45.080 and then the system generates a minute by minute sentiment 43 00:02:45.360 --> 00:02:49.159 for each company which is useful in different trading situations. 44 00:02:50.439 --> 00:02:53.680 And we have also another platform which is an internal 45 00:02:53.680 --> 00:02:57.960 platform for research. Is like a set of research tools 46 00:02:58.319 --> 00:03:02.840 we used to exchange in form about research. So what 47 00:03:02.919 --> 00:03:06.120 they have in common is the need to process real 48 00:03:06.240 --> 00:03:10.240 time let's say events, and to especially to self heal 49 00:03:11.039 --> 00:03:16.439 because we need we need to stay up. Obviously it's 50 00:03:16.439 --> 00:03:19.639 not one hundred percent, but we need to stay to 51 00:03:19.719 --> 00:03:23.879 be reactive to any kind of anomaly, and to. 52 00:03:23.840 --> 00:03:28.919 Process UH trades as far as fast as we can. 53 00:03:29.439 --> 00:03:29.759 Nice. 54 00:03:29.919 --> 00:03:32.919 So you mentioned in there like the need to leverage 55 00:03:32.960 --> 00:03:38.680 other libraries or languages outside of UH you know, Corelixer, 56 00:03:39.120 --> 00:03:42.120 and it seems like in today's development environment that seems 57 00:03:42.120 --> 00:03:45.599 to be increasingly the case because you have, like you'd mentioned, 58 00:03:45.680 --> 00:03:48.439 data scientists who are using their own tool set of tools, 59 00:03:48.960 --> 00:03:50.840 and they already have these libraries that are being built 60 00:03:50.879 --> 00:03:53.120 up by a whole community of people, and a. 61 00:03:53.039 --> 00:03:54.240 Lot of those tend to be in Python. 62 00:03:54.639 --> 00:03:58.919 And historically, just as web developers, I think we you know, 63 00:03:58.960 --> 00:04:01.680 you're you're accustomed to having work with multiple languages because 64 00:04:01.680 --> 00:04:03.639 you have a front end which might be written in 65 00:04:03.719 --> 00:04:06.000 JavaScript and a back end which is written in something else. 66 00:04:07.319 --> 00:04:10.120 Some people do approach, you know, take Node as a 67 00:04:10.159 --> 00:04:12.199 back end, but you know we have a lixer, or 68 00:04:12.199 --> 00:04:14.680 you've done Ruby on rails or Java or something else, 69 00:04:15.479 --> 00:04:17.399 and then you have mobile on the front end. You know, 70 00:04:17.439 --> 00:04:21.319 you're dealing with you know, mobile frameworks and mobile languages 71 00:04:21.360 --> 00:04:23.279 and different platforms. So it's like we are kind of 72 00:04:23.279 --> 00:04:25.720 in this area and this environment where we have to 73 00:04:25.720 --> 00:04:31.800 deal with multiple types of tools, multiple communities. And so 74 00:04:32.000 --> 00:04:33.800 one of the things I think is interesting is this 75 00:04:33.879 --> 00:04:37.319 whole idea about how Python has a lot of these 76 00:04:37.360 --> 00:04:43.120 libraries for doing like image recognition, for doing you know, 77 00:04:43.160 --> 00:04:45.120 like where you know, be it Python or not. But 78 00:04:45.439 --> 00:04:51.040 I might be using GPUs to do accelerated computing. And 79 00:04:52.360 --> 00:04:55.360 so what I loved about your article is that you're 80 00:04:55.560 --> 00:04:58.959 you're totally acknowledging that, right that there is stuff that 81 00:04:59.000 --> 00:05:02.959 we need to use that's not an elixir. So we 82 00:05:03.199 --> 00:05:06.480 want to basically try to leverage the benefits of each 83 00:05:06.800 --> 00:05:10.399 the strength of each language and each platform to do 84 00:05:10.439 --> 00:05:12.839 what they're best at. So maybe you can tell us 85 00:05:12.839 --> 00:05:15.839 a little bit about this blog post that you wrote 86 00:05:15.920 --> 00:05:17.639 and kind of give us an introduction to this topic 87 00:05:17.680 --> 00:05:18.680 of what you're trying to do. 88 00:05:19.720 --> 00:05:25.759 Yes, sure, the so the blog post is about real 89 00:05:25.879 --> 00:05:30.959 time object detection uh and the idea of using real 90 00:05:31.000 --> 00:05:35.680 time object detection in Phoenix UH. Well, in Elixir and 91 00:05:35.720 --> 00:05:38.839 then using Phoenix to render the result. 92 00:05:40.319 --> 00:05:43.279 So obviously the idea. 93 00:05:43.279 --> 00:05:45.759 Of the of the blog post is not to reinvent 94 00:05:45.839 --> 00:05:52.199 the will is actually to use something that is already there. So, 95 00:05:52.720 --> 00:05:55.920 as you said, there are great frameworks in Python and 96 00:05:56.720 --> 00:06:01.399 UH is there are also high level libraries like a 97 00:06:01.519 --> 00:06:09.560 civy lip which uses open cv under the hood, and 98 00:06:09.839 --> 00:06:13.000 UH it's super easy to use. And you can just 99 00:06:13.079 --> 00:06:16.720 pass an image to this to a function and UH 100 00:06:16.920 --> 00:06:22.839 dysfunction returns uh list of detected objects and UH and 101 00:06:22.879 --> 00:06:24.800 the coordinates of these objects. 102 00:06:25.040 --> 00:06:28.480 So in the in the article, I actually uh uh 103 00:06:28.920 --> 00:06:30.639 go through UH the. 104 00:06:30.560 --> 00:06:34.600 Idea of interfacing and trying to take advantage of this 105 00:06:34.800 --> 00:06:42.959 library UH inside the Elixir, and I went through different ways 106 00:06:43.000 --> 00:06:49.079 of communicating with with Python. Well, the simplest one is 107 00:06:49.120 --> 00:06:54.759 to just launch just run UH the Python process. The 108 00:06:54.879 --> 00:07:01.600 problem of this is that two for example, in this 109 00:07:01.839 --> 00:07:06.319 specific case, we have a neural network, which is a 110 00:07:06.600 --> 00:07:11.040 neural network, well actually a convolutional neural network, and to 111 00:07:11.240 --> 00:07:15.839 load this model it takes around two three second on 112 00:07:16.000 --> 00:07:20.399 a matwork pro and obviously we can't run this UH 113 00:07:20.560 --> 00:07:24.839 this script for each single frame or each single image 114 00:07:24.879 --> 00:07:29.000 we want. And so we need a long running process. 115 00:07:29.360 --> 00:07:36.079 That means we we keep the model in memory and 116 00:07:36.079 --> 00:07:39.560 and we keep the Python process in memory, and we 117 00:07:39.720 --> 00:07:43.240 control it via something is called a port. So a 118 00:07:43.399 --> 00:07:46.680 port in a in Elixir and in Airline, it's a 119 00:07:46.879 --> 00:07:53.600 mechanism to launch an operating system process which is external 120 00:07:53.639 --> 00:07:59.800 to the bm UH. And this brings a lot of advantages. Well, first, 121 00:08:00.079 --> 00:08:04.120 if the process crashes, it doesn't bring down the bim UH, 122 00:08:04.240 --> 00:08:05.519 so it's something external. 123 00:08:05.759 --> 00:08:08.680 And we can communicate. 124 00:08:08.160 --> 00:08:13.639 With this process via message passing and the standard IU UH. 125 00:08:13.720 --> 00:08:16.279 And what I do through the article is to. 126 00:08:18.079 --> 00:08:23.879 Is to use ports to build UH, to build UH, 127 00:08:24.079 --> 00:08:28.560 this communication to H with the with this UH Python 128 00:08:28.639 --> 00:08:34.240 framework and be able to link the camera from UH 129 00:08:34.480 --> 00:08:38.279 from the web browser, take the frames from the web browser, 130 00:08:38.559 --> 00:08:43.360 send this frame across UH a web socket to a 131 00:08:43.360 --> 00:08:49.200 Phoenix channel, and then Elixir sends UH these frames to 132 00:08:49.759 --> 00:08:54.759 the Python side detected which actually runs the the real detection. 133 00:08:55.639 --> 00:09:00.679 And and yeah, the the thing is is so easy 134 00:09:00.799 --> 00:09:06.519 to use this approach to bring and to use to 135 00:09:06.559 --> 00:09:12.440 take advantage of machine learning. And since Python is so so. 136 00:09:12.559 --> 00:09:14.360 Far on the machine learning side. 137 00:09:14.399 --> 00:09:18.440 Actually, one of the reasons Python has a huge adoption 138 00:09:18.600 --> 00:09:23.799 rate is because all these frameworks in the last ten years, 139 00:09:23.840 --> 00:09:28.039 the adoption rate increased a lot. But because these frameworks, 140 00:09:28.919 --> 00:09:34.519 not because the language itself. And yeah to me, to me, 141 00:09:35.759 --> 00:09:40.159 I think Elixir would be great. I actually would prefer 142 00:09:40.440 --> 00:09:47.720 to control a machine learning library with Elixir rather than 143 00:09:47.799 --> 00:09:53.240 Python with pipes and struct and like generate a struct 144 00:09:53.360 --> 00:09:55.440 with it, which is like a job. 145 00:09:55.519 --> 00:09:57.600 I want to run a machine learning job. I want 146 00:09:57.639 --> 00:09:59.039 to run and then. 147 00:10:00.759 --> 00:10:03.960 And then send this job to to a Python to 148 00:10:04.000 --> 00:10:08.240 a Python process with the with depart or with other 149 00:10:08.919 --> 00:10:10.200 solutions like. 150 00:10:10.360 --> 00:10:16.600 Uh, there are other ways like uh E p M D. 151 00:10:17.279 --> 00:10:21.519 So the is said there is a library called remember 152 00:10:21.600 --> 00:10:27.080 correctly pyre lank UH and UH, the it's possible to 153 00:10:28.039 --> 00:10:31.679 uh to create a Python node. UH, just to say 154 00:10:31.720 --> 00:10:35.600 there are other ways we can interface with the with 155 00:10:35.679 --> 00:10:37.000 the Python process. 156 00:10:37.159 --> 00:10:39.639 So one of the ways if I can just kind 157 00:10:39.639 --> 00:10:42.200 of clarify the way I understand ports to work. 158 00:10:42.759 --> 00:10:42.879 UH. 159 00:10:43.080 --> 00:10:46.399 I have used ports before, and UH in my case 160 00:10:46.440 --> 00:10:49.799 I was talking to we needed some you know, we're 161 00:10:49.799 --> 00:10:53.679 migrating a Ruby on Rail's monolith to Elixir, and so 162 00:10:53.720 --> 00:10:55.480 there's some things that Ruby still. 163 00:10:55.240 --> 00:10:57.960 Did best because of libraries or whatever. 164 00:10:58.320 --> 00:10:58.440 You know. 165 00:10:58.440 --> 00:11:00.360 I don't want to re implement an entire life library 166 00:11:00.440 --> 00:11:02.960 just so I can have this little piece of functionality 167 00:11:04.399 --> 00:11:07.080 so as ports as I understand it, Like what I'm 168 00:11:07.080 --> 00:11:10.879 able to do is I can start up the process, 169 00:11:11.200 --> 00:11:13.840 the OS level process that is running Ruby in my 170 00:11:13.919 --> 00:11:16.799 case or Python in your case. And so you're like 171 00:11:16.840 --> 00:11:20.399 you're describing you like this two to three second load 172 00:11:20.440 --> 00:11:23.399 time just to load everything that it needs for the 173 00:11:23.960 --> 00:11:26.679 for doing the memory yea, for doing its processing. And 174 00:11:26.720 --> 00:11:29.679 so once that load is done, then it's just basically 175 00:11:29.919 --> 00:11:34.519 it's messaging through like I think it's even just standard 176 00:11:34.759 --> 00:11:38.799 out kind of just directly kind of talking through the 177 00:11:38.799 --> 00:11:41.440 OS to the other process, not even through like an 178 00:11:41.519 --> 00:11:45.879 HTTP kind of you know port like you know web 179 00:11:45.960 --> 00:11:47.559 ports or anything like that. It's not a web servers, 180 00:11:47.559 --> 00:11:51.240 just talking directly to it. So that has a cost, 181 00:11:51.399 --> 00:11:54.279 I guess in terms of serialization about how do I 182 00:11:54.320 --> 00:11:56.600 actually communicate and how do I serialize my data to 183 00:11:56.639 --> 00:12:00.159 send it across to this other process. But that that's 184 00:12:00.159 --> 00:12:03.399 one of the benefits is uh, it's it can now 185 00:12:03.440 --> 00:12:07.919 be supervised by the beam, right, so if something goes 186 00:12:07.960 --> 00:12:12.759 wrong in my library and it crashes, elector can respond 187 00:12:12.799 --> 00:12:14.039 to that, right have you? 188 00:12:14.120 --> 00:12:16.559 Is that something you've been able to see as well. 189 00:12:17.600 --> 00:12:22.679 Yes, absolutely, this is exactly one of the advantages and 190 00:12:23.120 --> 00:12:29.080 uh uh with part, especially if we wrap a part, 191 00:12:29.519 --> 00:12:32.320 if we wrap a gen server around the part, it 192 00:12:32.480 --> 00:12:36.399 becomes super easy to uh to supervise it. 193 00:12:36.720 --> 00:12:39.639 So yes, in the article I, what I do. 194 00:12:39.720 --> 00:12:44.799 Is to uh wrap dispart uh with with the gen 195 00:12:44.879 --> 00:12:48.879 server UH, and it's possible to run like to start 196 00:12:49.360 --> 00:12:53.240 monitoring this part. So when the process crashes, we just 197 00:12:53.360 --> 00:12:58.080 receive a message to the owner process which is an 198 00:12:58.080 --> 00:13:03.080 exit message. And so in this case we can deal 199 00:13:03.360 --> 00:13:06.879 inside the gen server UH. And we can just like 200 00:13:07.080 --> 00:13:14.240 rerun the part without letting the supervisor do that, without 201 00:13:15.000 --> 00:13:17.799 having to delegate this to the supervisor. Or we can 202 00:13:17.960 --> 00:13:22.919 just decide to not handle the exit process the exit message, 203 00:13:23.120 --> 00:13:27.960 and what happens is that the the gen server process 204 00:13:28.039 --> 00:13:33.440 just crashes and if it's supervised uh, it's restarted again. 205 00:13:33.600 --> 00:13:36.799 Another another advantage is. 206 00:13:38.399 --> 00:13:42.960 Is that of this wrapping wrapping the part with the 207 00:13:43.039 --> 00:13:47.159 gen server is that we can only the owner, so 208 00:13:47.799 --> 00:13:52.279 only the process that starts the part can communicate with 209 00:13:52.320 --> 00:13:56.200 the part. So can send a message which is propagated 210 00:13:56.440 --> 00:14:01.080 via via standard input to the to the Python process 211 00:14:01.519 --> 00:14:05.480 and the same as receiving messages from the Python process, 212 00:14:05.639 --> 00:14:09.799 So only the owner can do this messaging. So if 213 00:14:09.840 --> 00:14:14.240 we want to use a port and let's click, it's 214 00:14:14.279 --> 00:14:17.360 not sure. But if we want to use this sport 215 00:14:17.600 --> 00:14:22.360 and send messages to this sport from many processes, we 216 00:14:22.399 --> 00:14:25.080 need a gen server that is the owner and takes 217 00:14:25.360 --> 00:14:29.519 the ownership of this port. So I think most of 218 00:14:29.600 --> 00:14:32.759 the cases is better to use a gen server, I mean, 219 00:14:32.960 --> 00:14:37.120 especially in this case. And also another advantage is that 220 00:14:37.440 --> 00:14:40.320 if like in this case, if we want to scale, 221 00:14:40.759 --> 00:14:45.120 if we want to scale horizontally, if a port is 222 00:14:45.279 --> 00:14:52.919 embedded is handled via gen server, is easy to spown 223 00:14:52.960 --> 00:14:55.840 a pool of distributed. 224 00:14:56.720 --> 00:14:58.200 Gen server workers. 225 00:14:59.399 --> 00:15:01.720 Yeah, I just want to point out how kind of 226 00:15:01.759 --> 00:15:06.200 incredible that is, and how Elixer's flexibility around like interrupt 227 00:15:06.240 --> 00:15:09.159 in this case with Python is so powerful because what 228 00:15:09.200 --> 00:15:12.440 you're doing is you're leveraging like the concurrency and the 229 00:15:12.480 --> 00:15:15.679 ability to manage processes that Alixir gives us and still 230 00:15:15.720 --> 00:15:19.639 getting like that powerful bit of Python functionality that is 231 00:15:19.679 --> 00:15:23.159 best suited to do that image detection. And that's why 232 00:15:23.159 --> 00:15:25.120 I kind of think it's fun to say, like, oh, 233 00:15:25.200 --> 00:15:27.919 Alixer is kind of good at everything, because in a way. 234 00:15:27.960 --> 00:15:30.519 It is because it's good at using the things that 235 00:15:30.600 --> 00:15:33.840 are better at the thing it can't do. And that's 236 00:15:33.879 --> 00:15:34.480 really awesome. 237 00:15:35.440 --> 00:15:40.960 Yeah, it's it's a fantastic glue in these cases. Is well. 238 00:15:41.039 --> 00:15:42.639 First is I. 239 00:15:42.480 --> 00:15:46.240 Think that the language itself is great, So most of 240 00:15:46.279 --> 00:15:49.000 the time I prefer to handle these kind of things 241 00:15:49.200 --> 00:15:54.039 and to glue things with Elixir. And at the beginning, 242 00:15:54.320 --> 00:15:59.639 I thought that this part mechanism could have huge limits. 243 00:16:00.200 --> 00:16:03.639 At the beginning, I thought, okay, if I send, because 244 00:16:03.919 --> 00:16:07.399 the experiment I did well. One of the experiments I 245 00:16:07.399 --> 00:16:12.559 did is to was to send the row frames like 246 00:16:13.360 --> 00:16:17.720 at thirty frame per second, so each frame is around 247 00:16:17.799 --> 00:16:23.720 the three megabyte, so I was sending around one hundred 248 00:16:23.799 --> 00:16:29.879 megabyte per second, and I thought, okay, maybe maybe I 249 00:16:29.919 --> 00:16:34.840 could have huge delays because there is messaging, there is IO. 250 00:16:35.799 --> 00:16:42.000 But actually each single message was sent the time the 251 00:16:42.080 --> 00:16:48.360 roundtree time I measured is like half millisecond for each 252 00:16:48.559 --> 00:16:52.519 single frame, which is I mean nothing, especially if we 253 00:16:52.559 --> 00:16:58.159 want to run a real time object detection at thirty 254 00:16:58.320 --> 00:17:03.960 frame per second. I mean, I was really impressed by 255 00:17:04.880 --> 00:17:10.000 how fast is this mechanism? And yeah, how easy is 256 00:17:10.079 --> 00:17:17.119 to take this approach and scale arizontally, especially when when 257 00:17:17.759 --> 00:17:22.240 when the party is embeddeding engine server, then we can 258 00:17:22.759 --> 00:17:25.880 use all the advantages of elix heir all the libraries 259 00:17:26.039 --> 00:17:31.880 like like a poor Boy or other libraries and supervision 260 00:17:32.039 --> 00:17:35.920 and it's I mean almost we get everything for free. 261 00:17:36.160 --> 00:17:36.359 Yeah. 262 00:17:36.400 --> 00:17:39.039 I love how Sophie kind of described that too, because 263 00:17:39.920 --> 00:17:43.079 I do think that is a strength of Elixir and 264 00:17:43.359 --> 00:17:46.640 the Beam in particular. It's just to say that it 265 00:17:46.759 --> 00:17:52.519 is a great interoperability messaging kind of platform, so that 266 00:17:52.720 --> 00:17:55.240 it's like it is the way I can manage to say, 267 00:17:55.279 --> 00:17:57.720 these are the things I want done, and then I 268 00:17:57.759 --> 00:17:59.920 can push that out through you in a distributed way 269 00:18:00.240 --> 00:18:02.920 out to the workers who this one might be rust 270 00:18:03.119 --> 00:18:05.160 and it's doing something, and this one might be Python 271 00:18:05.200 --> 00:18:07.920 and just doing something, and then bring all those back 272 00:18:08.359 --> 00:18:12.079 and be able to aggregate and report and do everything 273 00:18:12.119 --> 00:18:14.359 like that. So it really kind of comes back to 274 00:18:14.400 --> 00:18:16.279 it's like wow, you know, it's kind of like this 275 00:18:16.640 --> 00:18:20.359 telephone communication kind of system, and it's like, you know, 276 00:18:20.400 --> 00:18:21.839 it's like wow, who to thunk? 277 00:18:22.720 --> 00:18:23.880 So it's kind of fun. 278 00:18:23.920 --> 00:18:28.680 But what I love is like in my case, except 279 00:18:28.680 --> 00:18:31.160 for like a couple examples like I gave with like Ruby, 280 00:18:31.359 --> 00:18:33.440 where I was needing to go to something because of 281 00:18:33.519 --> 00:18:38.279 a library as a gem that I needed to support 282 00:18:38.440 --> 00:18:41.160 for Otherwise, it's like, I don't. I don't have any 283 00:18:41.200 --> 00:18:43.000 personal needs where I need to go out to those 284 00:18:43.039 --> 00:18:46.319 other things. I don't have performance requirements where I say 285 00:18:46.359 --> 00:18:47.359 I need to go to RUST. 286 00:18:47.680 --> 00:18:47.960 I don't. 287 00:18:48.000 --> 00:18:51.160 I'm not doing anything with right now with machine learning. 288 00:18:51.519 --> 00:18:54.119 We'd like to at some point in the future, but 289 00:18:54.240 --> 00:18:55.960 you know, at some point that's where I will go 290 00:18:56.119 --> 00:18:59.240 and I will be interoperability kind of talking with Python. 291 00:18:59.759 --> 00:19:01.519 So I love that I can do it, but I 292 00:19:01.559 --> 00:19:04.359 don't have to do it. Yeah, Elixer still does everything 293 00:19:04.400 --> 00:19:06.039 I need to write where I am right now. 294 00:19:06.599 --> 00:19:06.799 Yeah. 295 00:19:06.799 --> 00:19:09.759 I think it's also it's also kind of like a 296 00:19:10.839 --> 00:19:14.440 useful case to make or argument to have in your 297 00:19:14.440 --> 00:19:17.480 back pocket when you're dealing with adoption. You know, if 298 00:19:17.519 --> 00:19:21.079 you're trying to convince your team or colleagues to use 299 00:19:21.119 --> 00:19:24.160 Elixir for a particular project, and you know, people are 300 00:19:24.160 --> 00:19:26.160 looking down the line and they're thinking, well, what about 301 00:19:26.160 --> 00:19:29.079 when we need something that Python is better suited for, 302 00:19:29.279 --> 00:19:31.000 or what about when we need to solve this problem 303 00:19:31.079 --> 00:19:33.200 that you know, typically folks may reach for rest for 304 00:19:34.920 --> 00:19:36.759 you can kind of trop this out and say, like, 305 00:19:36.799 --> 00:19:39.079 we can use Elixir now so that we can move fast, 306 00:19:39.079 --> 00:19:40.759 so that we can leverage like a nice design, so 307 00:19:40.759 --> 00:19:43.559 that we can take advantage of all the fantastic Elixer 308 00:19:43.599 --> 00:19:46.680 things around concurrency and fault tolerance. And when we need 309 00:19:46.839 --> 00:19:49.039 something else, we need a tool that's better suited for it. 310 00:19:49.240 --> 00:19:50.799 We have it backed ourselves into a corner. 311 00:19:50.880 --> 00:19:52.640 We're not in a place where we need to, you know, 312 00:19:52.680 --> 00:19:55.359 spin up a separate service and another language and framework 313 00:19:55.480 --> 00:20:00.000 and then figure out what this communication mechanism would look like. 314 00:20:00.039 --> 00:20:04.319 I think one of the things that you mention in 315 00:20:04.359 --> 00:20:05.960 your article obvious is. 316 00:20:05.880 --> 00:20:10.119 That you know, you could have reached for HTTP communication. 317 00:20:10.240 --> 00:20:12.240 You know, both some API end points where you're going 318 00:20:12.319 --> 00:20:16.240 to communicate between your alexro Phoenix app and you know 319 00:20:16.279 --> 00:20:18.279 this this Python application. 320 00:20:18.440 --> 00:20:20.119 But we don't have to. We don't need that. 321 00:20:20.119 --> 00:20:23.400 Overhead, we don't need that complexity. Ports are one way 322 00:20:23.759 --> 00:20:26.519 among some others that Alexa just makes it super super 323 00:20:26.559 --> 00:20:28.960 easy to enact that kind of communication. 324 00:20:29.599 --> 00:20:33.599 Yes, compared to for example, HTTP, which is obviously the first, 325 00:20:33.680 --> 00:20:36.240 I mean the first thing that came to my mind 326 00:20:36.319 --> 00:20:39.119 when I wanted to do something like this, And the 327 00:20:39.160 --> 00:20:43.759 problem is supervision and especially on embedded devices. I mean, 328 00:20:43.799 --> 00:20:47.359 if I want to run this thing, a real real 329 00:20:47.440 --> 00:20:52.079 time outject detection system, that is that does something if 330 00:20:52.079 --> 00:20:55.920 I find some objects in the pictures in the frames 331 00:20:55.960 --> 00:21:00.799 coming from the camera, I have an if I run, 332 00:21:00.880 --> 00:21:05.079 for example, a Python worker that has flask like an 333 00:21:05.200 --> 00:21:09.920 HTTP UH server, I mean usually most of the time, 334 00:21:10.000 --> 00:21:13.960 especially on an embedded device, I want to supervise this 335 00:21:14.200 --> 00:21:20.079 process or this worker, and if I'm able to supervise 336 00:21:20.160 --> 00:21:24.720 it via elkx heir, I'm much happier, especially because this 337 00:21:24.839 --> 00:21:30.480 means I don't have to well, I know what's happening 338 00:21:31.160 --> 00:21:36.000 inside Elixir, so inside the Elixir, I know if my 339 00:21:36.359 --> 00:21:39.160 workers are fine or not, and I don't have to 340 00:21:39.200 --> 00:21:43.359 install some other services like well, I don't know kubernatists. 341 00:21:43.359 --> 00:21:48.200 Maybe not in an embedded device, but it's something that 342 00:21:48.279 --> 00:21:51.440 I could run on a on a different in a 343 00:21:51.480 --> 00:21:55.519 different case if I want to UH scale easily and 344 00:21:55.839 --> 00:21:58.000 supervise different services. 345 00:21:59.359 --> 00:22:05.839 So this parts and makes this much easier. 346 00:22:06.519 --> 00:22:10.759 And another solution, as Mark said, UH with with another 347 00:22:10.880 --> 00:22:15.519 solution to communicate UH for example, like with the with rust, 348 00:22:16.200 --> 00:22:19.759 it would be uh niffs. So I went through also 349 00:22:19.880 --> 00:22:24.680 that path and so yallow which is uh. 350 00:22:24.799 --> 00:22:28.079 It's an acronym for you only look once. Which is 351 00:22:28.079 --> 00:22:32.279 the algorithm behind the object real time object detection. This 352 00:22:32.400 --> 00:22:36.359 is a state of the art object detection system which 353 00:22:36.400 --> 00:22:41.000 is with a real really uh simple, with the n 354 00:22:41.039 --> 00:22:45.759 Vida GPU with its not cheap but around the five 355 00:22:45.799 --> 00:22:50.839 hundred dollars GPU. Uh, you can detect objects at a 356 00:22:50.920 --> 00:22:52.039 thirty different per second. 357 00:22:52.720 --> 00:22:53.279 Uh. 358 00:22:53.359 --> 00:22:55.799 And the thing is uh. 359 00:22:56.000 --> 00:23:02.039 The the creator of this algorithm they he did the 360 00:23:02.119 --> 00:23:04.759 C version C and COUDA versions. So I said, okay, 361 00:23:05.519 --> 00:23:08.759 maybe I could do a NITH. So first it's so 362 00:23:09.079 --> 00:23:13.559 far easier to uh do to use the part with 363 00:23:13.680 --> 00:23:17.160 a Python a high level library, then create an diffan 364 00:23:17.279 --> 00:23:21.440 see and trying to interface with to to bind to 365 00:23:21.480 --> 00:23:26.799 create a binding to this library and the But the 366 00:23:26.839 --> 00:23:31.119 main thing, Uh, what happened is that the first time 367 00:23:31.160 --> 00:23:36.240 I compiled there was an error in my ce neth 368 00:23:36.839 --> 00:23:41.519 and obviously uh it crashed. The problem is that everything crashed, 369 00:23:41.680 --> 00:23:47.079 the beam crashed, and so yeah, there are ways to 370 00:23:47.119 --> 00:23:49.240 deal with it. Well, first is use rust. 371 00:23:49.720 --> 00:23:53.200 The thing is, uh, I didn't find anything. 372 00:23:53.279 --> 00:23:57.799 I mean, there were some libraries, some bindings with open 373 00:23:57.920 --> 00:24:03.599 CV with rust. By the things when quite quite complicated, 374 00:24:03.759 --> 00:24:08.039 quite easily, so I said, okay, ports is far better 375 00:24:08.119 --> 00:24:08.319 to me. 376 00:24:08.400 --> 00:24:11.440 I mean, I'm not. I tried many things. 377 00:24:11.200 --> 00:24:16.359 But if I, if I would go back, I would 378 00:24:16.440 --> 00:24:19.279 just use ports in this case. In this case because 379 00:24:19.519 --> 00:24:23.319 the detection time is still around a twenty million second, 380 00:24:23.359 --> 00:24:30.400 which is reasonable, reasonable time to wait to delegate a 381 00:24:30.480 --> 00:24:33.920 job to a Python process. Obviously, if I just want 382 00:24:34.000 --> 00:24:37.599 to do a sum from for two floats, a part 383 00:24:37.960 --> 00:24:43.480 doesn't isn't isn't isn't the best way because just this 384 00:24:43.680 --> 00:24:50.680 messaging it costs the time. But and in that case 385 00:24:50.720 --> 00:24:54.519 maybe it's better to use to use a NIF. But 386 00:24:56.079 --> 00:25:01.039 for things that are longer that than you milli seconds 387 00:25:01.880 --> 00:25:09.119 parts is easier and safer uh and uh safer than 388 00:25:09.200 --> 00:25:15.880 cm in uh and um yeah so and uh and 389 00:25:15.559 --> 00:25:19.759 and also is we were able to just use a 390 00:25:19.839 --> 00:25:22.960 high level library rather than something. 391 00:25:24.440 --> 00:25:25.599 With a lower level. 392 00:25:25.880 --> 00:25:28.400 I think it's funny the library called Yolo. 393 00:25:30.160 --> 00:25:33.200 You know, I I thought it was like you only 394 00:25:33.240 --> 00:25:35.440 live once, but I can't. 395 00:25:35.480 --> 00:25:37.359 That can't be it. So that's really cool. It's like 396 00:25:37.359 --> 00:25:38.119 you only look once. 397 00:25:38.759 --> 00:25:38.880 Uh. 398 00:25:39.319 --> 00:25:41.000 But what I love about your article, and I just 399 00:25:41.000 --> 00:25:42.920 want to encourage you, dear listener, to check out the 400 00:25:43.000 --> 00:25:46.119 article because if you if this is an interesting topic 401 00:25:46.160 --> 00:25:48.240