video_id
stringlengths 11
11
| title
stringlengths 0
100
| text
stringlengths 513
648
| start_timestamp
stringlengths 8
8
| end_timestamp
stringlengths 8
8
| start_second
stringlengths 1
5
| end_second
stringlengths 2
5
| url
stringlengths 48
52
| thumbnail
stringlengths 0
52
|
|---|---|---|---|---|---|---|---|---|
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | parallel bayonet this work as a standalone auto regressive model and still be deployed on an Android phone hybrid models with much weaker or aggressive structure but that can be trained on a large escape could be revisited and and of course all these architectural innovations that help in long-range dependencies would always help in you know as you keep moving to bigger image this or a video or something like that these kind of ideas should up a lot so like a summary of auto regressive model could be that it | 00:11:02 | 00:11:38 | 662 | 698 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=662s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | is an active topic but a lot of cutting-edge to us and there's a lot of moscow for a new engineering and creative architecture design and larger models and data sets are clearly needed to you know realize the full potential of these class of models and standalone they are very successful across all modalities without any conditioning information like class labels so that's that's like a very appealing property of these models every Universal in that sense and also they can work without much engineering for sampling time so | 00:11:38 | 00:12:14 | 698 | 734 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=698s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | that makes them really look creative but but but nevertheless for production if you you should really cut down on the sounding time to be useful and so innovating on the low-level primitives was very important so that said there are a lot of negatives for aggressive modeling one is you don't extract any representation there is no bottleneck structure and sampling times not good for deployment it's not particularly usable for downstream tasks like for instance a language Maru you need to sample multiple times to see coherent samples | 00:12:14 | 00:12:52 | 734 | 772 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=734s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | so you can't just roll out a language model that's a software and there are no interpolations that you can see to visualize what the models actually learning and every time you sample it's going to take a long time to produce like a diverse set of samples so that's it about auto regressive models now let's look at flow models in flow models it all started with the nice architecture by loaned in and those the model was already producing very good digits on the endless data set and on the T of tedious it was producing | 00:12:52 | 00:13:27 | 772 | 807 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=772s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | reasonable phases but it really was bad on see far and SPH India said the samples were very blurry but it all improved with the real end we'd be architecture which introduced other kinds of flows and rational room to make the models better and then the glow model from King model was published where the real and Ruby model was taken to another level by making it prettiest much larger images and overdone in our lab called flow pass class advanced the likelihood scores for flow based models to competitive scores that with that of | 00:13:27 | 00:14:05 | 807 | 845 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=807s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | autoregressive models for the first time and this was done by this architecture engineering and scale so this shows the power of flow models of potential they have in terms of closing the gap in density estimation between autoregressive models without having the powerful or aggressive structure but at the same time being really fast with sampling and also potentially useful for inference so given all these practices there's a lot of future work left in terms of how to learn the masks how do you actually completely close the gap | 00:14:05 | 00:14:38 | 845 | 878 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=845s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | with our regressive models whether you want to use very expressive fluids but very few or whether you want to use shallow flows which are not particularly expressive but then keep on stacking them so that you can get a very expressive compose model how do you use multi scale losses for a trait and how do you trade off between your density estimates and your sample quality and how to use the representations you derive at various levels of the flow model for downstream tasks all these are like fundamental advances think about | 00:14:38 | 00:15:15 | 878 | 915 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=878s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | for flow models and also how do how do you carefully initialize so that flow models can train very fast so in terms of core achievements that you can aim for you can aim for producing low level samples which are truer models that have way fewer parameters the globe uses half a billion parameters for all the celebrity faces and that's unlikely a scale and how do you make it work potentially for even larger images how do you do dimensionality reduction with flows and think about other other flow models like conditional flow models and | 00:15:15 | 00:15:55 | 915 | 955 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=915s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | you know how do you actually close the cap and sample quality de Gans and also close the likely skoros gap between autoregressive models so the models would provide the pathway to do both and it's it's interesting to think about how to do all these things together so the negative of flow model says you expect to have the same dimension at every layer every stack of the flow and so it's unlikely to scale if your data is getting bigger and higher dimensional and unless you innovate on how to do dimensional reduction sauce it's unlike it'd be | 00:15:55 | 00:16:29 | 955 | 989 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=955s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | useful and you really need to carefully initialize and use things like AK norm for good numbers so that's that's another negative because it may not be directly usable for another modality or another data set or another kind of architecture so let's look at late engraver models will see the various different be strengths and weaknesses and what have been some visible successes in bas it all started with the original Emnes modeling by dirk Kingma where you could see various types of digits and strokes and the slopes of the strokes | 00:16:29 | 00:17:11 | 989 | 1031 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=989s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | and shades across multiple digits and then it got extended to much better more powerful data sets like Elsa in bedrooms by pix ovae and also image not 64 by 64 creating much better global sound globally more coherent samples 10 pixel CNN because of modeling latent structure and then there's the latent variable models innovation in terms of using hierarchical models and multi stack using hierarchical Laden inference and producing really high quality sound really faces on par with slow models so there are well-known applications of V | 00:17:11 | 00:17:55 | 1031 | 1075 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1031s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | like sketch iron and role models and BW is used for modeling visual concepts and there are applications like deep mines jeredy cry networks which does view synthesis of a separate view by taking in two provided views and embedding into a latent rifle and interpolating the lane space for a query view across across multiple possibilities and therefore you can just collect data in a completely new environment from first-person vision you can you can keep a track of all their poses when you're recording things and then in principle | 00:17:55 | 00:18:34 | 1075 | 1114 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1075s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | you could figure out how a particular scene looks like from any other viewpoint and therefore reconstruct the entire room or entire environment completely through this kind of a synthesis model that has rational inference so we have practically used in these kind of architectures and there are lots of advantages of EA's you get a compressed bottleneck representation you can get approximate density estimates you can interpolate and visualize what the model learns you can potentially get disentangle representations where | 00:18:34 | 00:19:07 | 1114 | 1147 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1114s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | different readings correspond to different aspects of data and it is like a model that allows you to do all these things together at once like you basically can sample so you are a gyrator model you have a density estimate so you can use for our distribution detection as a density model you have latent variables so you you do representation learning and you also have a bottleneck representation so you are able to reduce the dimensionality of your original data set so a VA is the only model that lets you do all these four things together and | 00:19:07 | 00:19:39 | 1147 | 1179 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1147s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | that makes it very appealing that said there are disadvantages you often end up with Lurie samples and assumption of a factorize Gaussian for the posterior or for the decoder this may be very limiting and you need more powerful decoders or more powerful posteriors and large scale successes are still yet to be shown and even though people have tried to like get more interpretable more disentangling variables by prioritizing the KL term over the reconstruction term the last it's still only work on toy problems and they may | 00:19:39 | 00:20:14 | 1179 | 1214 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1179s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | actually be better ways to do representation learning or generation or like yeah interpolation in some form hierarchical Layton's individually so expecting for one model to all of them well may be truly hard and so a we may not be the state-of-the-art models on anything but maybe a model that lets you do all all that it all these things recently well in using a single single single modeling framework so that's that that's the that's what you lose when you want is everything within one model so that these are the disadvantages to me | 00:20:14 | 00:20:55 | 1214 | 1255 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1214s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | but there's obviously scope for future work you can but you can use bigger decoders more powerful posteriors you can think about how to do hierarchical Leyton's to learn covers and fine-grained features and discrete Leyton's like weak uva and also large scale training like slow models have been done like glow or focus bus so next let's cover implicit models but we look at general adversarial networks and just just basically what what's happening ganz though we also covered moment matching energy based models in | 00:20:55 | 00:21:35 | 1255 | 1295 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1255s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | class the Gann samples the quality of Gann samples has dramatically advanced from the primitive samples that you saw in the original Gann where you saw X reasonably looking good faces but then the c4 samples it's not pretty cooing too critical in terms of what is the object or class of C far that's been captured but it certainly looked different from Larry BAE samples at the time next you saw DC Gann which clearly advanced some the some quality of dance to a state where again to assign you a looking much and much more exciting than | 00:21:35 | 00:22:18 | 1295 | 1338 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1295s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | any other model because the samples were much sharper and all these bedrooms were very high dimensional and then recently again giving again has been taken over by began stag and classic models were clearly careful attention to detail in terms of architecture design and also really really large-scale training like large pad sizes and a lot of stabilization tricks can produce these amazing photorealistic samples that you've already seen plenty of times in the class so I'm not going to go over them in terms of future work for Ganz I | 00:22:18 | 00:22:56 | 1338 | 1376 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1338s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | think I think it's really hard to bet against cans to say hey this is work cans weakened its most likely that if you put sufficient effort in engineering you can get it again to function well on those things as well but but nevertheless there's still more progress we made an unconditional cans more collapse and also more complex scenes and video generation will be cool for instance will be nice to get a model that works on real driving data where and a lot of pedestrians are walking and then you want to be able to simulate future | 00:22:56 | 00:23:30 | 1376 | 1410 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1376s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | you have to keep track of multiple people multiple objects multiple cars road signs and so forth so it's a very complicated jeredy modeling problem and it'll be interesting to see it ganz which are known to identify only a few cues in your dataset would they still work in such complex settings where you need to keep track of multiple things at once so future work in terms of modeling you can like think of more purchasable Lipsius knows better conditioning tricks like how to feed noise if your various levels like for instance stai again basically | 00:23:30 | 00:24:08 | 1410 | 1448 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1410s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | innovated their batch or instance normalization of how to design better architectures working on sampling and down something ops to use how do you how to do channels of sampling and done something without introducing a lot of parameters what is the right objective function for your discriminator and how to scale and train ganz in a stable manner for like larger problems and how to preserve it at various different levels like how do I instance noise a feature noise so that it can stabilize the training of the discriminator much | 00:24:08 | 00:24:41 | 1448 | 1481 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1448s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | better so all those things are very very interesting and think about in terms of negatives again scans if one could say there's plenty of engineering details and it's hard to clearly identify which is the most important core component that helps you reproduce these high-quality images and it's also very time consuming to ablate for these details so and and and and and it's very clear we need to improve on the sample diversity but then we also don't have very good metrics for evaluations so we need to work with what we have and even | 00:24:41 | 00:25:19 | 1481 | 1519 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1481s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | though it may seem like we're improving a lot on the current metrics we use for again evaluations objectively the sample diversity is not a spurious likelihood based models so how do we actually come up with better valuation measures also one thing to think about with all these aspects like good evaluations good metrics relations these are not particularly specific to the scans these can be said for any any any kind of model as with any other model so if you were to make a choice between Ganon or density model | 00:25:19 | 00:25:55 | 1519 | 1555 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1519s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | one would imagine you need a lot of engineering details for Ganz but it's not particularly true even for density models the architectural engineering has been comparable level of detail and you know trickery that you need for Ganz and secondly there is a lot of attempted theoretically understanding Ganz so the trade-off between having blurry samples versus of being okay with mode collapse is basically the same trade-off that you make when you care more about compression at the cost of sample quality was this you wanting to have | 00:25:55 | 00:26:34 | 1555 | 1594 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1555s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | really good samples at the cost of missing some modes so it's basically which direction of kale that you care about and the reverse direction you care about more if you don't want any spurious sample but the forward direction you care about more if you really want to make sure that your modeling is good and you're not going to make any mistakes even though your you're not gonna miss out anything you in there you may make some mistakes at some of some of the points so mostly apart from the fact that they can produce amazing samples cans are | 00:26:34 | 00:27:09 | 1594 | 1629 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1594s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | popular because they can work with much less compute for instance in order to generate a 1 megapixel image for an auto regressive model or even a Leighton space our aggressive model you need to use at least 512 course or TPU to do that because you need such large pad sizes whereas for gans you can make it work with a single V 100 GPU and then so there so that's that's one reason why gangs are clearly preferred over than 10 C models because I'm amount of time taking the train as a sample and you can also see better interpolations and | 00:27:09 | 00:27:46 | 1629 | 1666 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1629s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | better conditional generation in cans so this dis leads to adoption by people who are more interested in art and fine tuning to like interesting artistic datasets you're not particularly machine learning relevant and that's one of the other reasons again a speaker plot so on the bright side we can think about how like many technological advances have been possible without the correct science and so ganz can we consider in that way as well and this is a slide from young Conan the epistemology of deep learning where explains that | 00:27:46 | 00:28:22 | 1666 | 1702 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1666s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | several technologies in the past have preceded their science that explains them for example the steam engine was before the thermodynamics so it's doing better theory for ganz is something that could still be innovated on in the future so here is a taxonomy of generative models from in Goodfellas new ribs tutorial apart from Markov chain Boltzmann machines and Markov change and are the stochastic networks we have pretty much covered everything else we've covered Nate may fix Lauren and how do you exchange of variables scale the flow | 00:28:22 | 00:29:00 | 1702 | 1740 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1702s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | models or really be models all these are explicit density models and then we also covered approximate in steam models vary from our encoders the variation lower bound and then recovered implicit density model estate they can other models that I'm not being covered are not particularly popular or very used so that's the reason we focus on the more popular ones and if you have if you're if you have been and trained density models and you're figuring out which density model you should be using here are some pointers if you only care | 00:29:00 | 00:29:33 | 1740 | 1773 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1740s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | about the density estimates disco for our aggressive models you don't worry about sampling time here if you care a lot about sampling times in autoregressive may still be fine if your sequences are not that big or if you use lightweight models but if you really cannot afford to wait for the sampling time you really want really fast samples but you still want to go for a density modeling you could think about using Vikan regressive models like paralytic so CNN and you could also think of doing latent space modeling like like latent | 00:29:33 | 00:30:05 | 1773 | 1805 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1773s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | space or like a week you BAE you may probably not even needed quantization bottleneck it could still work with like continuous values and so models are also pretty billing for modeling continuous value data that density estimates for continuous value data especially even when they're actually continuous and it's hard to figure out how to even quantize them so so that that's that's another interesting aspect of flow models and if you also want to think about how how to have like representations and also sampling but | 00:30:05 | 00:30:46 | 1805 | 1846 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1805s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | you want to have a simple possible model v's with factorize decoders maybe the natural drugs so given given these appealing properties or density models like when would you use cans you would use guns when you really care about having good samples and you have really really large images high-quality images for and you don't want something photorealistic you have a lot of conditioning information like pose or the class or edge edge maps and you just want to add texture to them cans are really good in these initial image | 00:30:46 | 00:31:17 | 1846 | 1877 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1846s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | translation problems or rear the video and if all they care about is perceptual quality and controllable generation and you don't have a lot of compute this is often the case for any any kind of start up again it's like the best choice to go for so that's it for generative models next let's look at South provides representation learning it which is our final topic so south supervised image classification has seen rapid advance in the last one and a half years just the end of 2018 the top one accuracy of image net linear classification | 00:31:17 | 00:31:57 | 1877 | 1917 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1877s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | benchmark was 48 percent and now it's seventy six point five percent so this rapid advance has been made in multiple labs because of this mode of learning called contrast to learning and contrast the learning task can be simply summarize this a dictionary lookup task and there are two ways to do this pretext contrasted learning which is you either build it as a predictive coding task or you build it as an instance discrimination task and in predictive coding you have multiple mechanisms to do that once you either used end-to-end | 00:31:57 | 00:32:31 | 1917 | 1951 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1917s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | mechanism or you use the momentum decoder momentum encoder make using the momentum encoder for the keys and the predictive coding success story has been achieved in the contrast operator coding or CPC particularly the CPC version two and and and the instances combination success has been achieved in moko and sim clear moko means momentum contrast and Sinclair's into an instance contrast they use the corresponding mechanisms of contrast learning so let's look at CPC version two moko and Sinclair in terms of their positives and the negatives so | 00:32:31 | 00:33:12 | 1951 | 1992 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1951s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | CPC version two we're doing spatial contrast prediction so that principle is very generic and it can apply to any morality or domain so you don't need to know the underlying data augmentation in variances in this work and it can be considered as latent space channel tomorrow and also it's much easier to adapt for audio video text and perform multimodal training disadvantages it splits your input into a lot of patches or frames or even audio chunks and therefore your inputs are now your inputs are now basically split into | 00:33:12 | 00:33:48 | 1992 | 2028 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=1992s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | a lot of different parts that you have to carefully delineate and you also need to carefully pick what part are you predicting from what so that involves a lot of design choices to make type of parameters that you can only know by trial and error so that makes it really hard for you to use it on a domain or task that you don't really understand well and then you require multiple forward passes for these smaller versions of the inputs now and so that means that you be pre-training on something much smaller but potentially | 00:33:48 | 00:34:19 | 2028 | 2059 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2028s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | fine-tuning are much larger versions of the sequences or images so this may not be an optimal thing to do when you're doing local predictions local spatial predictions Bosch num is hard to use so applying mass ROM is hard but then you really want to use batch room for a downstream task so that makes CPC version too little sore in sense it's not particularly suitable for downstream tasks if you really care about state-of-the-art performance and finally the splitting process mechanism is very slow on a on a matrix multiplication | 00:34:19 | 00:34:55 | 2059 | 2095 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2059s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | specialized hardware like GPUs so it's because you do a lot of reshapes and transposes and so it's never an optimal thing to do so here's the summary of moco one of the main advantages of moco is it is very minimal so it's very easy to use and replicate and it has no architectural change can be easily applied for downstream tasks there is no notion of a patch and it's distilling in variances for images using data augmentations and so the pre-training procedure looks very much like supervised learning and therefore it can | 00:34:55 | 00:35:31 | 2095 | 2131 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2095s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | get comparable or even better results and the momentum encoder memory bank can assume adds a lot of stability to the training and decouples back size from the number of negatives and therefore this lets you train with way fewer GPUs than what's needed for CPC or like methods the disadvantage with moco is that because you introduce momentum and date you need to figure out what's the right decay rate for that and that has an extra type of parameter and another disadvantage is in image augmentation the invariances may not be applicable to | 00:35:31 | 00:36:05 | 2131 | 2165 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2131s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | other modalities so this may be in method this works only for a visual image recognition and finally let's look at simply er which can be considered as an end-to-end version of Tomoko where you just look you're using all the negatives from your batch and there is no momentum encoder so advantages or sim clear are the same as that of moko with the additional advantage that you don't have a momentum in kora now so it's going to be asked minimally supervised learning but the disadvantage is now you just need really large batch sizes | 00:36:05 | 00:36:39 | 2165 | 2199 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2165s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | because you need a lot of negatives because moko decouples the negatives from a bad size it doesn't need as much compute as sim cleared us and similar to moko they documentation invariance may be very specific to image recognition so in terms of future work left for sauce provision the gap between some supervised learning and supervised learning is to not close if you consider just the same amount of compute training time and the same candidate augmentations use so and also fine-tuning to downstream tasks the | 00:36:39 | 00:37:13 | 2199 | 2233 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2199s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | gains are not significantly high enough that the paradigm shift has been made in vision so that way maybe new objectives are also needed and finally all these sub supervised successes have relied on using image net and it's not clear if supervised learning we just work from images in the wild or from the internet which is really the dream and which is really why people wanna do something so that's it for like subspace learning as in in terms of utility for downstream tasks let's look at always learning in the context of | 00:37:13 | 00:37:54 | 2233 | 2274 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2233s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | intelligence like being able to act in an environment so here is a video of this quake3 game where yeah like that you can see some characters and then you can see some bullets that there are going to be fired and you know you see all these different walls and fires and other characters and when you're looking at all this you're able to already accurately parse the scene make sense of what's going on and you're also able to clearly separate out the objects from what's not objects and and so we need to be able to do that as well we shouldn't | 00:37:54 | 00:38:36 | 2274 | 2316 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2274s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | be working at the level of pixels we should be able to predict the future in a much more semantically in space and so modeling the pixel space for these high dimensional videos is really hard and in order to build really dungeon agents which that can planning faster than real time we should be able to do it in the lane space that's more abstract so how do we do that what is the right kind of abstraction to build and how do we learn role models in that Lane space that can this ignore noise and work in a much more semantic space it's really the | 00:38:36 | 00:39:09 | 2316 | 2349 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2316s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | hardest question to think about and this has also been summarized multiple times by omnicom that if you have very good internal world model you'll be able to plan with it and a wide lot of mistakes there and our relation usually makes and and how to do that is one of the most important questions so if you want to have the overall view of subspace learning across all these different problems for image recognition we saw or assesses like city scene workers in clear moco version to transfer learning it works really well in language but the | 00:39:09 | 00:39:48 | 2349 | 2388 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2349s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | exact details will be covered in a future lecture and transfer learning and vision also works reasonably well now been shown in CPC and moco but there's like close to nothing in terms of how to use of supervised learning for RL so that's the very ripe area for future and then as far as like you know using sound supervision in the context of general intelligence is considered its it's potentially going to be extremely useful in the context of transfer learning and learning use of abstractions for planning or imaginations so that's just a lot of | 00:39:48 | 00:40:26 | 2388 | 2426 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2388s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | work to be done there so that's that's that's it for the summary of the class it's pretty much ends with our original motivation which is how do we build this intelligence cake and and a lot of it is gonna be done through supervised learning and and so in terms of future lectures they're gonna look at more applied topics which are not falling into the main main main lecture stream which is that we be looking at semi spread learning we'll also be looking at the whole area of one square learning for language which is language models | 00:40:26 | 00:41:08 | 2426 | 2468 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2426s | |
1sJuWg5dULg | L8 Round-up of Strengths and Weaknesses of Unsupervised Learning Methods -- UC Berkeley SP20 | and bird and then finally we look at how representation learning or supervised learning has been applied in the context of reinforcement learning so and and we will also cover things like how to do unsupervised distribution alignment that is given completely to different data sets with a lot of common information how do we align the two manifolds together and without any prior data and you see how generative models and unsupervised learning can be used in the context of building compression algorithms so that's the next next | 00:41:08 | 00:41:41 | 2468 | 2501 | https://www.youtube.com/watch?v=1sJuWg5dULg&t=2468s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | [Music] okay so hi I'm Lucas easy one first of all thanks to debug and to be a developers for inviting us today so what we will try to do today is share some of our experiences when it comes to IP architectures we'll also try to give some recommendations about what to do and what not to do when you're designing an architecture in a larger enterprise environment we will try to give examples where we can from our own experience and you will see we will cover actually several topics so feel free to ask any questions if you know we go cross one | 00:00:00 | 00:01:11 | 0 | 71 | https://www.youtube.com/watch?v=OOMFR6snocY&t=0s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | topic too fast because number of topics that we want to share so okay let's start so actually what is an enterprise IT landscape and how does it look like you probably guess that it's big given and I we work in telcos most of our careers which is a really big act landscape but just to make it more clear I'd like to compare it with the buildings and let's say a neighborhood so it looks like something like this you know it's a being it's modern it's neat it's beautiful I'm kidding of course it doesn't look like that at all probably | 00:01:11 | 00:02:06 | 71 | 126 | https://www.youtube.com/watch?v=OOMFR6snocY&t=71s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | only chief architects thing that looks like this actually it looks something more like this or this or this if you're lucky the reason is that most enterprises have been building their IT for the last twenty thirty years and doing it mostly by adding new systems on top of existing ones rarely decommissioning your ones so now you have a really big combination of really cool and modern stuff and some really uncooled and all that stuff that's it that actually makes it really interesting but also challenging so but | 00:02:06 | 00:02:53 | 126 | 173 | https://www.youtube.com/watch?v=OOMFR6snocY&t=126s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | what happens at one point since you're building building building on the other it starts to crumble and you actually have two choices you need to you need to change something or as you like to call it these days do a transformation project and then you can do it in two ways basically you can go all the Intuit Big Bang changed 80% of your landscape or you can do it progressively in our experience I have done one transformation that was a big bang approach let's actually first don't don't do Big Bang transformation | 00:02:53 | 00:03:40 | 173 | 220 | https://www.youtube.com/watch?v=OOMFR6snocY&t=173s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | projects try to do it one step at a time and progressively if possible so how can you do it well first you need to identify what are your biggest pain points what is the what is the thing that is most troubling your business with the current IT landscape then think of a better way to doing that when you do that take a step back because you're probably already going to white and you're looking at the project of three to four years you can't do all of it do a minimum of what you can think think of but but it makes sense that it really | 00:03:40 | 00:04:25 | 220 | 265 | https://www.youtube.com/watch?v=OOMFR6snocY&t=220s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | solves the biggest issue that you can and that you can integrated with the existing systems that you have so in order for you to do that there are actually three things that you need to consider and have in mind for us in our experience microservices architecture is a must in doing that why not just because it's a hype and everyone in doing that but because you can also release in a lot of books that talk about macro services architecture it's as people say so are done right it's finally finally we have the technology and the means to do | 00:04:25 | 00:05:10 | 265 | 310 | https://www.youtube.com/watch?v=OOMFR6snocY&t=265s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | service-oriented architecture in a good way to make it scalable and maintainable so that's why macro services architecture is a must but when you can really do a lot of wood or doing a lot of bed is a in integration part so you need to be smart when it comes to integration patterns and also is in all development unit who developed development practices system otherwise won't be able to tactical so those are actually free areas that we will try to cover in our presentation first the micro services architecture the main | 00:05:10 | 00:05:56 | 310 | 356 | https://www.youtube.com/watch?v=OOMFR6snocY&t=310s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | message that I want to convey here is the micro service isn't the support application now don't get me wrong I have nothing against people would I like it actually but though many times I've seen someone will the Java spring boot application deploy it on the docker container and says I have a macro service you don't have a micro service unless that micro service has data when we come to the definition of a micro service in its essence it's a small autonomous application that can handle one domain independently autonomous and | 00:05:56 | 00:06:34 | 356 | 394 | https://www.youtube.com/watch?v=OOMFR6snocY&t=356s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | independently you can't do that if you don't share data you can't do that if you're creating a facade on my or proxy my cursor that keeps baiting legacy and this query in that legacy for all day I mean you can do it and that there are uses for such such applications but you can forget about any kind of performance or scalability if you query an existing legacy banking systems for every time someone the queries your API on a micro service and needs to get some data that's the most important peak we will talk later | 00:06:34 | 00:07:15 | 394 | 435 | https://www.youtube.com/watch?v=OOMFR6snocY&t=394s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | even will talk about what are good ways to get data out of legacy in inter macro circuit so I'm going to that you'll hear that later other thing that is important is how you design your micro service what is important is a lot of times people have a set of functionalities you put it in a micro service and you you just go from there it's a great way to build a my product what that means is build a micro service that's actually monolith just your deploy it on docker and the quality Mac series in micro service design you should really look at | 00:07:15 | 00:07:57 | 435 | 477 | https://www.youtube.com/watch?v=OOMFR6snocY&t=435s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | domain driven design that's another concept you have great books about it especially the part you have a book by Eric Evans domain driven design when he talks to bounded context but what that means actually in a nutshell is that when you design a data model for an enterprise you can't design one data model that will work for the entire enterprise because that's its action is called an enterprise functions for example you take customer entity and you talk to finance you talk to say all the top customer customer care each of those | 00:07:57 | 00:08:46 | 477 | 526 | https://www.youtube.com/watch?v=OOMFR6snocY&t=477s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | Department look at the customer differently in finance they look at did he pay his bill and suffocated in sales they look at good services is he using can we upsell cross-sell upgraded in customer care they are looking at does have problems does he have problems in the past what kind of problems etc so the models are not different and what you need to do is you need to establish a boundary in which your model is unique and that is basically in one domain we will could have one model that is unique and then when you talk to another team or | 00:08:46 | 00:09:30 | 526 | 570 | https://www.youtube.com/watch?v=OOMFR6snocY&t=526s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | whatever the physic designing another domain you need to establish what the boundaries are exact like and what is important how to transform and connect your model with their model it's really a good concept I haven't talk any more about it but you should look at it really a lot of materials online about it so the next topic when it comes to micro services actually this is one of my favorites and also a lot of debate is around debt when it comes to the communication patterns between micro services how do they | 00:09:30 | 00:10:12 | 570 | 612 | https://www.youtube.com/watch?v=OOMFR6snocY&t=570s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | communicate with each other there's actually two big patterns one is choreography the other is orchestration choreography you have micro services that are independent they function as let's say individuals with a set of rules in which they talk to each other but there is no central micro service or whatever that is orchestrating them in orchestration you have a central entity that is orchestrate that is telling the other micro services what to do which one is better neither because it depends on the use case that you have | 00:10:12 | 00:11:03 | 612 | 663 | https://www.youtube.com/watch?v=OOMFR6snocY&t=612s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | mostly in whatever you are designing NIT there is no one-size-fits-all you need to be open-minded and you need to look at exact use case that you have so for example a choreography is a really great it is actually my preferred method because in choreography the main benefit is that these micro services are completely decoupled when one goes down it doesn't influence the rest if you need to remove one micro service and implement it in using a different technology or whatever you can do it because they communicate asynchronously | 00:11:03 | 00:11:55 | 663 | 715 | https://www.youtube.com/watch?v=OOMFR6snocY&t=663s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | they have a set of interfaces that are standardized and this really gives you a lot of flexibility but what is a problem with choreography is that if you have a complex system or a complex process that you need to implement it's really hard to visual the communication so if you can't visual that which visual visual visualize that communication don't use choreography you can visuals visualize it by using open tracing and technologies like that there there are ways but if you can't do it don't want choreography orchestration on the other | 00:11:55 | 00:12:39 | 715 | 759 | https://www.youtube.com/watch?v=OOMFR6snocY&t=715s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | part is great for example when I use it in an order execution process when you have a process that is sequential and you really need to follow the exact amount of steps that are needed to talking for example in telco when you do order execution you need to create a customer you need to create his assets you need to send his equipment to the delivery service you need to talk to other telephone operators regarding number transfer and stuff like that you need to activate the subscription for the customer in billing | 00:12:39 | 00:13:30 | 759 | 810 | https://www.youtube.com/watch?v=OOMFR6snocY&t=759s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | there's a lot of activities that you need to do all of those can be implemented as different micro-services but many work is straight you are really sure that everything is happening the very data truth but you shouldn't do orchestration in code because the same is choreography you won't have any visibility of that when you do orchestration my advice is use a BPM tool if a really good BPM tools on the market that you can use and they intend visibility in the control the process is great so just to show you this is one BPM tool and | 00:13:30 | 00:14:16 | 810 | 856 | https://www.youtube.com/watch?v=OOMFR6snocY&t=810s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | this is how it how it looks like you have where are the process instances some are in incidence some are not so it's a it's really a good way to see what is happening in that process you really do if something is amiss in incident you can easily see what's happening so the level of control here is really great so that's my advice when it comes to orchestration but the most important part is no one-size-fits-all so let's move on just some quick best practices when it comes to micro-services design actually the only | 00:14:16 | 00:15:13 | 856 | 913 | https://www.youtube.com/watch?v=OOMFR6snocY&t=856s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | don't here is don't put everything in code use the technologies that are on the market and that are really great for example if you have a lot of business rules that you need to implement there are great rule engines that are fast and that you can use instead of writing millions of ifs and cases in your code the only trick here is if you're losing in a row if you are using using and there are there two to two ways you can use the rule and you can embed it in your code and then it's very to just use the engine and edit code or you can | 00:15:13 | 00:15:52 | 913 | 952 | https://www.youtube.com/watch?v=OOMFR6snocY&t=913s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | build it as a independent standalone service but if you do that it's important that it's a passed rule that the rule engine only executes rules it doesn't query any database or back-end or anything for data it should look something like that this is request this is response you give in the the rule that it needs to execute you give him the data and then it executes the rule inhibit the response back and it should be in memory and it should be really fast as you see here it should really be a number of milliseconds the seconds that you can | 00:15:52 | 00:16:41 | 952 | 1001 | https://www.youtube.com/watch?v=OOMFR6snocY&t=952s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | count on your fingers and toes because if it's slower than that you can forget about any kind of performance that uses the rule engine so it's really important that it's fast proof it's in memory and it doesn't query anything it's just an engine also use artifact repositories they're really easy to use really beneficial use tools for logging don't log to file system database and then if you have an incident spend three hours wearing debt use pools the terror in on the market protect the self is for monitoring use time series databases | 00:16:41 | 00:17:26 | 1001 | 1046 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1001s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | because they can handle a lot of data and they have great visualization tools that you can look at in the operations part so I'm not going into specific technologies that you can use but for all of these technologies you have open source solutions if you want to know I can tell you what we are using but maybe later on or in the question space and the last but probably the most important things he is work together with business because a micro service or any softer that is its own purpose doesn't make sense so you should you should build it | 00:17:26 | 00:18:15 | 1046 | 1095 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1046s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | together with people from business and you should build stuff that really solves some business issue that can be some new capabilities of extra functionalities they can be just that you are doing refactoring transform suddenly some old implementation into micro service because you want better performance which will result in better customer experience these are all valid reasons but don't do it without talking to business and having really good memories because basically micro service should be analog to a business domain | 00:18:15 | 00:18:53 | 1095 | 1133 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1095s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | and it should implement business capabilities so that is really an important part here so they could actually be the first step when you're designing the micro-service know what is the challenge and what are the capabilities from the business side that you need to solve and then go to the technical part and that's it about the micro services architecture I will now give the work to even we will talk more about integration about integration patterns so when we were doing this kind of migrations basically integration | 00:18:53 | 00:19:31 | 1133 | 1171 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1133s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | become as you move to micro service integration becomes more of a hot topic than it was before so some of our experiences don't use multi layer of API clothes what I mean by this we have a situation currently were fronting is going bacon that is calling second packet that's calling terror that comes forth bacon so basically you have this connection of flows where each layer is introducing their own bottlenecks there are slowdowns so don't do this second is what Luke also mentioned in micro service architecture try to do the | 00:19:31 | 00:20:17 | 1171 | 1217 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1171s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | main driven design in telco industry where we are happy that we have a global guidelines regarding this there is a community gathered around TM for telemedicine forum organization that already built for us some models so we know what are the customers what are the products what are the services resources payments so we already have domain driven design that can be followed basically we already have domain prepared that just need adjustment for a specific tau and this domain is extensive enough so that you can extend it without let's say many | 00:20:17 | 00:21:02 | 1217 | 1262 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1217s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | travels of course no brainers don't use non scalable or low performance by non-scalable I mean try to not reach the limit where you need to pay for extra licenses or where you are on all the technology that you cannot horizontally scale I mean you can always add more CPU power but this is something that needs to be avoided and also try to use the try enough to use no documented API I know we have cases where there's an API working for 10 years it works good it's just that developers that did that long time goal of the company and you | 00:21:02 | 00:21:45 | 1262 | 1305 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1262s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | don't want to rely on such API because when the problem starts then you will need to redesign the API in a matter of hours just for it to work so if you have something that is old and I mean how could we have a lot of old technologies that are there because they work and we have systems that are old 10 or more years you have stuff that is not documented so what to do in integration patterns try to use data provider API so if you cannot come close to the data by varying the database try to come to the layer that is closest to that database | 00:21:45 | 00:22:26 | 1305 | 1346 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1305s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | layer in some cases you can of the show some on the shelf solution where it is a black box you only have the API provided by the solution no one is telling you ok in the transformation let's replace it you purchased if you want to have return of investment over it and basically you should use it but you don't want to use first API that is going second API that is calling this code safety I try to consume it directly and then use rule engine to apply any rules that you have over data to provide domain-driven the main data that these | 00:22:26 | 00:23:09 | 1346 | 1389 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1346s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | needs for that particular purpose and try to do something that is a try to ruin something that is easily upgradeable to new principles so if you have old APR don't let you see okay you need to spend a little more effort to be compliant with the new proposals new strategy invest in that API clean it out and upgraded don't built everything from scratch but of course when you have 10 or 20 years of development a lot of things are piled around so sometimes you just need to have a clean cut and do it from the beginning | 00:23:09 | 00:23:48 | 1389 | 1428 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1389s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | regarding the integration patterns we found let's say something like five turns I will present some of here and what is important is that one pattern doesn't fit all so first when we are trying to figure out what betters we first want split it up by CQRS principle when we looked at our logs and how users are accessing our data most of the data access is read access so a lot of people on the self-care device is going to see what is the state of their consumption are they paid have they do the bills and stuff like this so most of | 00:23:48 | 00:24:33 | 1428 | 1473 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1428s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | the data is coming from down code to the customer and then then you have a payment of bills activating some new adam strange's tariff and this is the smaller amount of command like queries so basically one pattern will not feel cold try to separate your integration patterns in your API by CQRS principle do stuff enhancing if if you have problems with executing the commands give the user the info okay it is your request has been received and do it quickly and then we're on if you have five ten or 20 seconds where you need to process this | 00:24:33 | 00:25:19 | 1473 | 1519 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1473s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | data is Luka mentioned our orders are complex sometimes we go from creating the customer to checking the dista to see if the speed is good enough so that he can receive some HD channels it takes some time so try to do it innocent way and try to keep it simple so these are some of the patterns I will share with you so this is the CDC of data replication better so we have legacy beckons legacy API what we are trying to do is copy the data from these databases to our application database and then do a data transformation from legacy model | 00:25:19 | 00:26:03 | 1519 | 1563 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1519s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | to new the main model so in legacy VIII you often have situations where in some applications there is a mix up of many models we have for example customer data and with five different databases so this shouldn't be the case in the domain driven design so what we are trying to do is pull this data as fast as possible through the data replication layer and then transform this data to a domain driven date and then when fronton describing your microservice you have data already prepared in the model that is by the domain driven design and what | 00:26:03 | 00:26:46 | 1563 | 1606 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1563s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | is important here data replication it's not only a CDC so depending on how and is the data change how often users needs to be aware of this data change it can be a job it can be a night and drop so basically it's up to you to see what fits your purpose and how you do the replication CDC he's only one of the options that that is applicable the second pattern that we saw is when you have off-the-shelf solution or where you need to consume something not directly from the database so how we did it for example in our case | 00:26:46 | 00:27:30 | 1606 | 1650 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1606s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | this is the users consumption API that is connected to a different billing systems and this data exchange really really frequently CDC and the amount of data that is coming through through the mediation system it's really big so CDC would just introduce the override that is not good enough so what we did first we simplify the architecture currently we have these multi layer API flows that we're introducing their timelines and bottlenecks now we connected either directly to the database or where we had all the shells who should be connected | 00:27:30 | 00:28:09 | 1650 | 1689 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1650s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | to a soft services that that of the shell solution presented and then we are doing the query this solutions transforming it in specific service layer and then presenting it to the customers of course I put the cash here because usually users don't want to see every second but what their data consumption so you can put this kind of data in a cache for 10 or 50 minutes 10 minutes should be enough for users to buy safe if he wants to return to the to the consumption data he can check it out but also need to be aware of you need to | 00:28:09 | 00:28:55 | 1689 | 1735 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1689s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | sometimes give him a pull to refresh capability if he really wants to see data exactly this time you can do it by demand but most of the users just when login to the South Korea they want to see the current consumption and if they jump over to the details you already have this data cache you can you can show it to them okay and the third layer I was mentioning regarding the commands that are being executed basically if you have situation where we are executing for example changing the death of a customer and this takes some time | 00:28:55 | 00:29:33 | 1735 | 1773 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1735s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | because a process needs to be run you need to connect to all those legacy systems that are responsible for changing the tariff from the what's a BSS layer to OSS there sometimes this takes time so trying to see what are the transactions that you have minimal number number of errors so if you have a change start process that is really if qualification of well staff user can jump into is really good at first then your execution part will mostly be successful but if it's slow you can use the pattern where you store the message | 00:29:33 | 00:30:11 | 1773 | 1811 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1773s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | you tell immediately to the user your request has been stored and then you start this in a synchronous way executing this message when this is successful send a push notification to the user so we have self care applications we have SMS these we can send to the user now your type has been activated so try to use whether you see that you have a really higher percentage of successful executions trying to do this in an ASIC way okay so other things regarding data aggregation and data consolidation don't do it on demand so | 00:30:11 | 00:30:52 | 1811 | 1852 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1811s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | this is more let's say applicable to beckoned for fronting players or front-end closer api's don't try to do on the on demand or on fly aggregation to Mapes try to set up a domain and prepare it how how easy it is fitting don't try to do aggregation over many domains it really means that you did something wrong if you need metrics if you need aggregated metrics try to prepare it in advance by relying on event-driven messages so instead of querying the database to see how much SIM cards customer has you can do once he is activating a new | 00:30:52 | 00:31:37 | 1852 | 1897 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1852s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | SIM card or something like this put this invent and store it somewhere in the bucket and then when you want to know how much SIM cards because queried bucket don't go to the database and try to query select counter from now number to return number of SIM cards also if you have situations where some of the micro services are sharing the same database but in a different schema don't try to consolidate data on the database level database should be treated as private fields so they are owned by microcircuits and there is a layer of a | 00:31:37 | 00:32:20 | 1897 | 1940 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1897s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | business logic above this data where perhaps when you are querying Sunday you won't receive a correct response so if you need to aggregate something prepare in advance do it in advance and don't do it on the fly especially directly on the database later it is okay as this is done by data where perhaps this is their job but I'm speaking about direct confirmation of data from digital channels from self-corrects also regarding the micro service approach and api's try to do some sort of management this really makes things easier so you | 00:32:20 | 00:33:00 | 1940 | 1980 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1940s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | can free API management tools you have a really expensive one so depending on your budget try to fit something but it is really important that over your api's you know who is technical and business owner you have the governance you know when new version of API will be set up you know when old version will be retired you will know who are in some cases we needed to introduce throttling to find out who are our consumers because some people we didn't even know we're consuming the PLC cool procedures started complaining about | 00:33:00 | 00:33:49 | 1980 | 2029 | https://www.youtube.com/watch?v=OOMFR6snocY&t=1980s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | timeouts so API management where you know what is the current version of API where we know who is consuming it and who is the owner of it this is a way to go don't treat ideas only as one protocol API so this is what is the domain and layer separation the main separation is domain driven design separation and layer separation is try to keep things controller service and repository to keep them out of each other way so every layer has its own boundary and the boundary shouldn't be crossed if you manage to do this then it's not a | 00:33:49 | 00:34:30 | 2029 | 2070 | https://www.youtube.com/watch?v=OOMFR6snocY&t=2029s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | problem for you to have an API that has repository and service and then it has controllers towards rest towards GRP see towards the Kafka or other streaming platforms so try to separate it both horizontally and vertically clean architecture programming the development practices so I try to split them up by organization level and T blow so in order to have a good development and good api's some rules in the organization needs to be set up it talking in small companies when you have a small team and startup you are | 00:34:30 | 00:35:13 | 2070 | 2113 | https://www.youtube.com/watch?v=OOMFR6snocY&t=2070s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | creating your own rules but on organization level you need to have support from operations department commits have support from DBA from system department so some rules needs to be set up what you don't want to do is to have multiple text acts try to find the text type that is okay for you and try to stick out with it you don't want to get yourself in a situation where some developer developed micro services in I don't know Godwin ergo because it was really right technology for him at that time and that feel of the company cool | 00:35:13 | 00:35:52 | 2113 | 2152 | https://www.youtube.com/watch?v=OOMFR6snocY&t=2113s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | govern disk so try to have a tech stack that is let's say a future proof and that you see it is really something used in beckons on beckons where let's say lucky here we are not as volatile eyes for example JavaScript frameworks became Java before we have Java now just we have different flavors but this is it also try not to limit the tech stack in a way that I don't know if you need something where you see graph databases will be good don't be limited by ok do it in the positives or something like this don't do like I say tech | 00:35:52 | 00:36:33 | 2152 | 2193 | https://www.youtube.com/watch?v=OOMFR6snocY&t=2152s | |
OOMFR6snocY | Building high performance and scalable architectures for enterprises—Luka Samaržija & Ivan Sokol | development I mean this is no transformation you are trying to move your team also from all these new technologies and you have a bunch of experts on all technology and their juniors in new technology and sometimes when you have deadlines the decision will be you we can do it in 30 minutes in all technology or in 3 hours in new technology invest in new first time it will be three hours next time it will be two and a half I mean no one no one no one learns King in 30 minutes so you need experience you don't want to go | 00:36:33 | 00:37:08 | 2193 | 2228 | https://www.youtube.com/watch?v=OOMFR6snocY&t=2193s |