AWS re:Invent 2019 review

AWS re:Invent 2019 review

Another year and another Cloud8 participation in Re:Invent. As always, we bring the news we find there from an analytical point of view and not just by listing topics and providing descriptions. To do this, simply access the Re:Invent summary – https://aws.amazon.com/blogs/aws/ . The idea here is to provide an analysis and tell you about curiosities and points of interest that can help the community and customers.

Previous years: 2018 , 2017 , 2016 , 2015 , 2014 and 2013 .

The full list of releases and technical details can be found on AWS website and YouTube channel.

Disclaimer : The comments below were created by Cloud8 and do not represent any position of AWS.

AWS re:Invent 2019 review

Strategic Summary

Re:Invent has been starting earlier and earlier. Two weeks before, the news blog is full of new features. Certainly an AWS tactic to increase awareness of the importance of the event and bring maximum visibility to you. The joke in the middle is that they monopolize so many technical advertising channels (not just the press, but individuals, social networks and the community), that the company that dares to make an announcement during this period will be lost and not very relevant. Anyway, at this point they have mastered the domain of attention.

The event begins on Sunday night with the reception known as Midnight Madness. Controversies aside about the good or bad taste of the ‘party’, the first news from within the event is announced. This year, at midnight sharp, Matt Wood came out to announce AWS Deep Composer (more about it in the next section) – apparently proprietary hardware has been successful and should be very profitable (see Echo Alexa, FireTV, Kindle, etc.) although this is very niche. The presentation of this keyboard coupled to a backend with Machine Learning reinforced the position of investing heavily in AI/ML and this was what was seen in Andy Jassy’s keynote.

On Monday night it’s Peter DeSantis’ turn to talk about global infrastructure. The focal point was to highlight how AWS has the capacity and performance of a commercial supercomputer. The analogy is simple: a supercomputer needs to have the architecture and be technically as performant as possible to run HPC (high performance computing) loads. If AWS meets the challenge of behaving like a supercomputer, it will handle virtually any type of load. And this is what he demonstrated. From the evolution of the network, which reaches 100 Gbps, with its own hypervisor known as Nitro and customized network protocol EFA, placement groups to guarantee logical and physical proximity of VMs with very low latency, network hardware to support thousands of nodes without degradation and without congestion and finally integrated HPC libraries. A cluster of new C5n instances (100 Gbps) was compared with a Cray supercomputer in a weather prediction simulation. Performance scalability was practically linear as nodes were added. As this was not the first time that AWS talked about weather forecasting, I believe that solving this use case was treated as a personal challenge internally…. cases of CFD use (fluid dynamics – Navier Stokes) and the case of Formula 1 (more below) were also shown.

DeSantis also talked about Machine Learning and that whenever we create a model we should think about separating types of instances for training and inference. For training, you need a lot of GPU and HPC, while inference uses another type of instance with a greater number of GPUs. The recommendation is to use p3 and p3n for training and g4 for inference – and there is now a new ‘inf1’ instance type of a GPU chip created by AWS. A typical case of ML on a gigantic scale is Alexa, which has the model trained periodically with clusters of p3n instances and then operates the requests (inferences) with g4 and inf1. A BERT training benchmark with optimized ML libraries was also shown, which are the best on the market.

Finally, and for the good of the planet, DeSantis confirmed that AWS has the objective of having 100% of the energy used come from renewable sources by 2030 and by 2040 being carbon free.

Keynote 1 – Andy Jassy

It is now customary for Andy Jassy to include in his keynote data on AWS’ dominance in the market, leadership in the Gartner quadrant, list of startup and enterprise customers, growth of the customer base and other indicators of extreme success. And, logically, emphasize the step of innovation and launch of new features (a disclaimer is worth mentioning here, as it would be necessary to filter “duplicate” launches such as qualifications in new regions, limit expansion features, etc. – logically without any demerit, but to guide the relevance of what is a mega launch – “New: Quantum Computing” – versus a practically irrelevant triviality – “ElastiCache supports 50 characters in the name”). Added to these facts, a very aggressive approach encouraging (healthily) migration to AWS. In previous years the discourse was that AWS is a reality (Cloud is the new normal), that innovating and making mistakes is cheap and that everyone should test the cloud. It’s not like that anymore. AWS no longer sees any reason to wait, with mature, stable and secure technology, it urges everyone to migrate as quickly as possible. Before, it was the developer, IT who led the Cloud adoption process and took the message about practicality and speed to the C-level. Now the change must come ‘top-down’ from senior leadership, with aggressive goals (he mentioned the case of GE, where the technical team had reservations about migration, but the CIO ordered it to migrate anyway. I don’t imagine that the audience technician was very happy to hear this…). Everything revolves around being fast: “Don’t let paralysis stop you before you start”. Lift and shift. “The hunger keeps on growing” (DMB). Rearchitecting is for later. And for all this to work well, he stressed that builders need to be trained. The forecast is that in a few years there will be a lack of up to 1 million technical professionals (sic) in the market to deal with cloud. The numbers shown that only 3% of workloads (USA) are in the cloud explains the call for migration. Still a lot of market.

He also talked a lot about Nitro (hypervisor). How it solved a range of technical challenges and enabled expanded innovation and product acceleration that would not have been possible before. Combined with the launch of new chips – Graviton 2 and the inference chip, computing on AWS has become more performant and cheaper. A significant increase in the performance/cost ratio is expected with the new types ‘m6g’, ‘r6g’, ‘c6g’ and ‘inf1’. They will be available for EC2, ECS, EKS and SageMaker.

There is always the traditional challenge to Oracle, even more so this year when Amazon finally migrated the latest Oracle database. Apparently AWS will miss having an ‘enemy’ and has already turned the cannon on IBM and Microsoft, provoking the migration of Mainframe and SQL Server (and Windows – more below). We hope to see jokes and hints at upcoming events with the new demonization of the mainframe and SQL Server.

Data – There is always news in relation to data. The name ‘Data Lake’ took a while to officially arrive on AWS and is now used in several contexts. S3, as always, continues to be expanded. Security scanning tools (S3 Access Advisor) and ease of access/security with S3 Access points have been added. Improvements in Redshift, Aurora, Athena (Federated Queries) and ElasticSearch in addition to the launch of managed Cassandra (more details in the technical section).

ML – The first release was Deep Composer and there were many others within the SageMaker suite. The highlight is, without a doubt, Sage Maker Studio. Basically a development interface (IDE) where you can create predictive models from a mere CSV file, or, logically, more complete/complex data. After the model is created, there are integrated debugging, monitoring and continuous improvement mechanisms. It generated a lot of curiosity among the conversations we held. From an applied ML point of view, CodeGuru was presented, which performs source code review, Contact Lens to transcribe and analyze Connect calls (Contact Center), fraud detection (Fraud Detector) and Amazon Kendra (search for documents within of the company considering natural language and generating references between them). A controversial point in relation to Artificial Intelligence (or rather Machine Learning) is that studies indicate that we may be reaching a valley of new things. The applicability scenarios of machine learning models should reach a ‘limit’ soon and new approaches beyond ML need to be pursued and be applicable. Several companies worldwide are chasing the new AI hype, but so far ML is the predominant one.

Hybrid / Private – Launch that I believe will be one of the biggest future drivers of revenue and growth: AWS Outposts + AWS Local Zones and Wavelength. Original in the 2018 announcement was to be the “Outpost for VMWare” and ended up being expanded to AWS Outposts and VMWare Outposts. The idea (although implicit) is to relaunch the datacenter and create a hybrid environment with a structure of physical hacks containing the main private AWS/VMWare stack and interconnected with the public cloud. Notice how the idea expanded. It started as a hybrid extension for companies that wanted to run VMWare and now it can be placed anywhere: companies, cities (welcome AWS Local Zone!), factories, countryside, etc. With the super appliance (pardon the simplification) AWS places “mini AWSs” wherever necessary, multiplying processing power, meeting private demands, reducing latency and integrating with 5G (Wavelength)! The first AWS Local Zone is in Los Angeles and Wavelength already has a partnership with Verizon to meet demands. One of the cases led by Volkswagen presented Industry 4.0 challenges, such as sensors, latencies, etc. An Outpost inside the factory would help solve challenges… Imagine the multiplicative power of this idea worldwide!

With all the releases and benefits, there was a negative aspect. Andy criticized Microsoft a lot because of the Windows licensing change . He rightly pointed out that this change is bad for customers and takes away freedom. In the end, he called for the end of Windows: “Closing Windows” and for everything to be migrated to Linux. What happened was not a ban but an increase in costs with the introduction of a new licensing model for dedicated servers – see Windows licensing FAQ on AWS . It’s bad, of course because it increases the cost, but there are no impediments. It is unnecessary to create a war. What did Steve Balmer himself say when he declared that Linux was a cancer, he increased the culture of aversion to Microsoft and in the end he himself already stated that Azure is mostly Linux . Another point that will be interesting to monitor is how Windows will support ARM technology (new Graviton chips) and whether AWS will support Windows in the new flavors. Waiting.

As you can see, Andy and DeSantis said almost everything. There is little left for Werner – although it is obviously a strategy to take the launches out of focus, as the area covered in all IT fields by AWS is already taken and mega launches tend to decrease (I hope I am mistaken and forgotten – Much of the fun of this market is following the launches and understanding how they fit into the ecosystem).

Werner’s phrase of the year was “There is no compression algorithm for experience” (self-explanatory). He focused much of the keynote on the Nitro hypervisor and how it enabled hardware innovation and the development of new products and improvements – new types of instances, Outposts, performance and security gains. The most interesting thing was to notice how the architecture of the infrastructure stack evolved – hardware and software. If, in the beginning, cloud was an assembly of components from various suppliers: processors, storage, network equipment, hypervisor, open source, etc., now with complete control of the technological chain a new era of optimization and efficiency gains has begun. The integration and fusion of many of these components to meet specific workloads is evident. Producing your own chip, your own hypervisor with built-in APIs, your own BIOS to manage security and unique scenarios (thousands of pseudo ENIs for microcontainers, EBS mounting redundancy, patching without reboot, etc.), network protocols for HPC (EFA ), even running a query in S3 is being done at the ‘edge’ (in this case within S3 without having to bring the data to the client, which is expensive and time-consuming), end-to-end productivity is gained and profitability is increased. Ultimately, in a gigantic topology, efficiency gains, even if marginal, make a huge difference.

Go build . Another of Werner’s phrases. It became a YouTube channel, where he travels the world following unique AWS application cases. Worth checking out.

Containers – Andy released Fargate for Kubernetes – ‘transformation’ from per-instance model to vcpu and memory model for Kubernetes, extending beyond ECS. The advantage is not having to manage and scale instances at the risk of underutilization and spending more. It is also a way to run ECS and managed Kubernetes, but with the trade off of being confined to the AWS model. Werner complemented the announcement with a demonstration of Fargate auto-scaling and how it compares to traditional EC2 auto scaling. If you take into account the response time of traditional auto-scaling, bootstrap, hooks, etc., and the microsecond initialization of microcontainers, Auto Scaling and Fargate’s response time are much more predictable and performant. It was a message to use Fargate.

One point that was not touched upon at the entire event, despite some speculation, was MultiCloud. While Google and Microsoft talk openly and amicably about competitors and integration, AWS doesn’t touch on the subject. It is natural for the leader to leave this point aside for now, but it will certainly have to be addressed one day – putting MultiCloud on the table would indicate that they are uncomfortable and would give space to competition, something that Amazon culture greatly rejects in our humble perception. .

Highlights and launches

Below are some of the most relevant and commented ones.

AWS Deep Composer

Another one for the series of ‘toys powered by ML’. Joins DeepLens and DeepCar. Composer is an electronic keyboard that is integrated with an ML backend. In the cloud it runs another ML algorithm, Generative AI. This model learns from data from different sources (text, image, music, etc.) and is capable of generating original content, as if using a little ‘creativity’ for AI (concepts like ‘intelligence’, ‘creativity’, etc. that are used in these areas are very difficult to define, beyond the limits and contours, but we will use the general idea….). It works with two neural networks communicating, where one generates and the other ‘corrects’, hundreds of thousands of times until a quality indicator is reached. For a more advanced view, I recommend the Cantor’s Paradise channel on Medium . In the demo, Matt Wood played a sonata on the keyboard and the deep composer completed it with other instruments – guitar, bass, drums – generating an orchestrated song. There was also a brief presentation by a professional musician who played a ‘4 hands’ composition (pardon the anthropomorphism). Not just about music, another example is the application in the arts (already mentioned in a past bulletin). See how interesting it is to apply the Van Gogh style learned in other works – and not to mention the Deep Fakes that use this technique…

EC2 Compute: Launch of new types. ‘inf1’ . These are chips with GPUs created by AWS itself for Machine Learning inference use cases. Types ‘m6g’, ‘r6g’ and ‘c6g’ were created on top of the new Graviton chip (ARM architecture) which has increased in number of supported cores and memory and can theoretically run any type of load;

Networking: increasingly complete and powerful VPC topology: AWS Transit Gateway – Multicast, Inter Region Peering, Network Manager, Accelerated Site to Site VPN. Multi-account and multi-region interconnection power;

S3 access advisor: S3 access analysis. Answers questions about who (identity, VPC, IP, etc.) has access to the buckets. With growth and complexity, it became more difficult to understand the topology and discover access permissions;

S3 Access points : following the Access advisor line, it seeks to simplify access to S3 objects by creating another layer with simplified permissions. Examples: configuring a bucket and certain objects for access by a VPC, IAM Role, etc. You can repeat access for as many objects as necessary. I imagine that behind the scenes, the access points are transformed into an IAM policy, which becomes huge and humanly unmanageable – hence the need to segment and simplify;

Federated queries for Redshift and Athena : perhaps one of the most useful and least hyped features of the entire event. AWS allows Redshift and Athena to make queries using other data sources (RDS and S3). Imagine you use Redshift to store historical data and Aurora for ‘hot’ data. With a single query and join you can bring all the data. Or save code tables in Aurora and not need to replicate elsewhere and so on. This functionality follows a principle that I heard in more than one session: AWS will not create a database that fits everything and prefers to use the best one for each scenario (RDBMS, key-value, timeseries, ledger, etc.), this way the need for integration in some way and use of data lakes are implemented by Federated;

ElasticSearch: One of the most common problems with ElasticSearch is the amount of used and unoptimized disk space. As the data history grows, the service starts to become too expensive and starts to require archiving maneuvers that take time and take away focus. AWS has developed a storage layer, which they are calling UltraWarm , which should significantly reduce this cost. In preview.

Cassandra : One of the few database ‘flavors’ that AWS did not yet manage. It is now added to the family of RDSs;

SageMaker Studio (Experiments, Debugger, ModelMonitor, AutoPilot): create predictive models in a (very) easy way. From a simple CSV file, or more complete/complex models, it generates a predictive analysis with the best tested algorithm. From the presentation we understand that the steps for this scenario would be something like: import the data, the studio understands the columnar model and the column with the expected results, tests 50 algorithms separating data for training and data for testing, presents the adherence result and you you can choose the model. Afterwards, it will monitor the inference and, if it deviates from a threshold, it will generate an alert so that it can be retrained. It seems very powerful to start with – you have to check, after the first step, what a more advanced evolution would look like. Build, train, tune, deploy.

Redshift with new type : a new type was launched – RE3 – which brings decoupling between CPU and storage – the current model has little flexibility between scaling CPU and storage, as when needing any of the resources you have to provision more instances, regardless of whether the other resource will be used. The cost is still quite high, so we should expect other types of instances in the future that are more affordable;

Lambda. Lambda has matured a lot over the year and has been released in small releases in recent weeks. One feature that caused celebration was the definition of competition . Pre-defined competition leaves a pre-heated pool to receive requests;

CodeGuru : code review and profiler using ML trained with data from Amazon itself. The idea is to scan the code base, and not just individual files, looking for bugs, improvements, possible competition problems, memory leaks, bank connections that remain open, etc. Conceptually, it is very interesting because of the scope coverage and being able to use the aggregated knowledge of a large base like Amazon and follows the line of other tools such as Github’s own Code Review and plugins in the most used IDEs. You can also run a profiler in daemon, which brings the most ‘expensive’ lines. The joke in some posts was to estimate the cost of reviewing the Linux kernel (despite the code being in C and Assembly and CodeGuru doesn’t support it)… check it out ;

Fraud Detector : fraud detection using ML with historical Amazon data. Talking to some fintechs that were at the event, there was no great excitement. The consensus is that fraud in Brazil has its peculiarities (like any other country) and that it is continuous work and immediate execution. This module would be another plugin and not a replacement for any existing or already implemented solution. To be seen how the adoption will go.

Contact Lens for Connect : transcribe and analyze Contact Center calls using ML. Implements a series of insights with sentiment analysis, trends in customer service, which ultimately helps to improve customer relationships. Support for the Portuguese language was not very clear;

Kendra : a search engine/aggregator of a company’s corporate documents. The difference is using ML to categorize and contextualize documents, linking similar subjects and generating a search portal. For those who remember Google’s yellow appliance launched in 2002, the idea is the same only with ML and in the cloud;

Security: a summary of the top 10 security topics by a solution architect;

Outposts : is the hybrid / private cloud version of AWS and VMWare with AWS products. Specific hardware connected to the public cloud and placed next to the customer. Already available in the US and other regions, but not expected to arrive in Brazil (see strategic discussion on OutPost + Local Zones). Notice how even though it is a product with extremely high added value, AWS was able to automate hiring to create scale! And ‘magically’ all the ecosystem’s tools and products already work with the endpoint pointed to an Outposts. “Cloud8 for Outposts” as soon as we have the first customer using it 🙂

Technical notes

There were more than 2,000 lectures. I comment here on what we managed to watch and the conversations we had:

Nitro : the ‘star’ of the convention. Lectures and highlights in the keynote. The first version emerged with the purchase of the Israeli company Annapurma and since then Nitro has evolved to be the fundamental piece of software for all AWS hardware. The achievements were numerous: offload of network I/O and disk I/O load, minimal virtualization and baremetal overhead (c5), security module and isolation of instruction execution between tenants, death of dom0 (who has already touched with Xen you know the risk of this guy…). Nitro has APIs where it is possible to control actions such as attach/detach disks in a much more efficient way. APIs decouple management and enable scaling. Microservices even within the hardware! Nitro has evolved to be embedded on a dedicated chip that runs all of these virtualization modules and can be patched without downtime. It enabled the creation of new types, a significant increase in performance and the creation of AWS Outposts;

Mainframe : elected one of AWS’s new enemies, just as it is/was with Oracle. In several possibilities there was talk about how to migrate loads. In the case of Goldman Sachs, the order was to ‘eat around the edges’, removing less critical workloads and creating a connector with the database. A fact not directly linked is the launch of the RDS Proxy which manages the connection pool (in addition to other aspects) and protects the database against a flood of requests, it has a lot of CICS and tuxedo vibes… (these are for the veterans) and it will certainly be a tool and an argument to accelerate the shutdown of mainframes;

Security : AWS showed a very interesting approach in the new way it identifies security configuration. Through Boolean modeling of all permissions (IAM) – codenamed Zelkova – and using a technique known as SMT Solver, it can create permitted paths and answer questions about who accesses what. There is an academic paper explaining how they did it. They did the same for S3 access;

Amplify : framework for creating application interfaces, has evolved a lot and is integrated with provisioning tools, such as SAM. There is a very interesting step-by-step guide on how to create a ‘modern’ application – Mystical Misfits . Who remembers the Pet Store?

Oracle : one of the lectures showed how thousands of Oracle bases were migrated. In the end, it wasn’t exactly a migration but rather a ‘refactoring’, where bases and systems were equally modified. In other words, it was not a pure de-para from base to base. Customers need to know that it is not as simple to migrate as it seems, depending on the resources that were used;

CDK : there was a lot of talk about CDK (Cloud Development Kit). Basically, it is a way of provisioning infrastructure using a programming language (Javascript, TypeScript, Python, Java and .NET) with all its advantages over templates: objects, modularization, sharing of a common code base. The CDK runs the code by transforming it into a CloudFormation template and then provisioning it. I suggest looking for Deep Dive sessions like DOP402 and the Workshop – https://cdkworkshop.com/ ;

API Gateway : a very important launch, but also not very celebrated, was the HTTP APIs feature. If you use API Gateway as a proxy for a Lambda, perhaps without the authentication, throttle, WAF features. etc., you are certainly spending too much. Either migrate to an Application Load Balancer, or now you can use HTTP APIs , which promises to reduce costs by 70%. Any lecture by Eric Johnson on this subject is worth checking out;

Event driven and microservices : lots of material about event-driven applications;

Amazon Detective : new product to help with forensic analysis – collects data from GuardDuty, VPC Flows and CloudTrail in a single interface and seeks to correlate events to help find security events;

Amazon Builders Library : the only release made by Werner, it is a collection of articles about how architectural challenges are solved by Amazon itself. In their jargon, “something you don’t learn at school”. It’s worth checking out.

Quantum Computing:

This topic is worth a dedicated section for the noise generated – it was the main “trending topic” on Twitter during re:Invent.

The first controversy was the launch made by the news blog, which immediately raised the question of why it was not launched during a keynote. Something so impactful should be in Andy or Werner’s keynote.

We attended the presentation of the product, led by Fernando Brandão – Brazilian, a professor at Caltech who participated in the project that led Google to declare quantum supremacy and which generated an immense debate about the feat. There was a conceptual theoretical introduction – qubits, technical challenges, Shor’s algorithms, applicability. Next, another AWS manager talked about productization. The idea is to provide ways to run algorithms designed for a qubit scenario. Algorithms that would run (much) more efficiently on quantum computers. To meet this demand, AWS acquired systems from 3 manufacturers: D-Wave, IonQ and Rigetti and will transform them into ‘coprocessors’ routing specific demands, instead of running on traditional instances. A collaborative initiative called “Quantum Compute Lab” was announced.

The answer to why it wasn’t released in the keynote became clear. It is something very, very new and is in full maturity. To give you an idea, Google’s Sycamore processor only has 54 qubits. For several practical problems this number has to rise to more than 100K (due to error correction and the challenges caused to physically stabilize the computer), in other words, we are far from that. It will certainly be very interesting to follow the future and see when AWS will launch its own quantum chip.

From a business point of view, the launch of this product makes AWS’s DNA very clear: surrounding absolutely everything that is being done in IT and leaving no room for competition. No room for hype. If someone has made some noise, or has something that is making money, AWS/Amazon will be there sooner or later.

Business cases:

F1: last year the AWS partnership for real-time data processing and AI to show statistics and probabilities of maneuvers had already been revealed (it already appears on TV broadcasts!). This year, Rob Smedley took on the challenge of making F1 more competitive in 2021. The idea is to create a car with an aerodynamic design that does not generate so much turbulence in the air after passing and harm those behind trying to overtake. Today, a car in the wake of another loses around 40% of its ‘downforce’ (the force that ‘sticks’ the car to the ground, unlike the wing of an airplane that propels it upwards) and creates a disadvantage when overtaking. With new CFD simulations carried out on AWS, the idea is that only a 7% loss will be achieved for those who are 0.5 seconds behind. There will be a jump in competitiveness ;

Goldman Sachs: the CIO, who is also a DJ, spoke about the new retail bank plus credit cards and the pivot of the bank that was previously investment-only. He particularly talked about how they are helping to move away from the mainframe;

BP: migrating all-in for cloud;

Insitro: ‘digital biology’ company that mathematically simulates DNA and helps find new drugs more easily. It has used supercomputing/HPC resources a lot;

Cerner: transformation in health care;

Avis: case about how the car fleet is connected and reporting problems, events and optimizing use;

Saildrone : this case is very interesting. Saildrone is an autonomous aquatic drone company whose mission is to map the planet’s oceans by collecting all types of data: salinity, temperature, pressure, sea currents, ocean depth, biomass (fish, plankton, microorganisms). With the data, you can improve weather forecasting and monitor global warming. Very impressive.

Ultimately, new opportunities were created and more challenges for everyone’s daily lives.

Count on Cloud8 to help you with this administrative and automation burden, as well as take advantage of the constant updating of our platform! There are many features that we can and will implement to further improve the platform.

Feel free to share this article ! Questions, criticisms or suggestions, get in touch.

Thank you

Renato Weiner
Cloud8 CIO / Founder

Share if you liked

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *