Event Location: Maydenbauer Center, Bellevue, WA
Approximate number of attendees: 300 – 500
Event Duration: 9 hours (8am to 5pm) + after party
- Talks were short and manageable in length (30 min each) so I could move around
- All topics were relevant and understandable
- I learned about cloud-edge computing and more about Kubernetes
- This was my favorite swag: Tech Tarot Cards
- Great dedication to diversity by giving out gender pronoun pins, providing diversity activities and events
- Excellent venue: lots of space, tables, electric outlets, and Wi-Fi
- Nothing. I fully enjoyed this event.
Would I go again?
Definitely. I recommend this to anyone wanting to get breadth on current cloud topics in the Seattle area.
Caution: This is a long one.
Fireside Chat with Mark Russinovich, CTO Microsoft Azure
- New Azure features announced recently: Azure IoT Edge, Data Lake integration, New China regions
- Azure IoT edge: a lot of computing will move out to the “edge” to optimize for bandwidth consumption and responsiveness of computed results.
- Github: Microsoft will not change what’s already there and fully invests in and supports the open source community
- How are people using blockchain in non cryptocurrency applications? Non-centralized security where multiple parties can see and verify a set of actions in a distributed networks.
- What is Quantum? It is a type of computer that can solve intractable problems in computer science that cannot be solved by a traditional computer.
- Is Microsoft planning to invest in quantum computing? Yes, Microsoft is getting ready for it along with an SDK and Q#, a language for quantum computing.
- Where do you think the best space to invest in technologically is? AI and ML – we’re at a breakthrough point where it’s still unknown how much we can do with it.
- Live poll: People are very unsure about what to think of Microsoft acquiring GitHub.
Did I like the talk?
No. This seemed like a giant advertisement for Microsoft and I could have done without. The comments on block-chain and quantum compute were interesting but very much sidelines.
3 Back-to-the-Future Trends in AI Systems – Carlos Guestrin, Apple and Amazon Professor Of Machine Learning at University of Washington
- We are in a period of resurgence in AI. Previously we saw a dip in its popularity.
- In the 80s, AI was primarily rule based systems that were not as useful or correct.
- Now, we are using machine learned models we are able to go into applications where we cannot enumerate all rules in spaces where there are a lot of possibilities. For example, we previously could beat humans in chess and now we are able to move to use Go without needing to describe the full space of the game.
- Spark uses large amounts of data with low compute per datum and massively scales out. We are moving towards Deep Learning where we have huge amounts of data (where huge > large) and there is high compute per datum that requires compute grids and node scalability.
- We moved from specialized hardware to specialized with Spark. Now with deep learning we need to move to more specialized hardware once again.
- AutoTVM: How do you make sure you choose the right coding patterns to optimize for your hardware when training models using a tensor language over different hardware? This project is meant to learn the right parameters to maximize use of hardware and optimize efficiency.
- We are trying to move to simpler adoption of machine learning. Currently you need a lot of expertise in order to understand data, use the tools, and manage the hardware needed for training.
- Two trends: pre-trained models where generic data is sufficient and the need for developers to create models for their specific scenario.
- Turi Create: allow developers to create their own domain specific data to easily train models via transfer learning from existing generic model training.
- Another trend is moving towards making the internal workings of machine learning and AI available to everyone instead of being a black box.
- Providing transparency into how a result was reached helps understand what features of the input are affecting the output as well as understanding where error might be coming from.
- The data you choose to use with define the user experience.
- Example: initially when film was being developed, they created chemical exposure guides that optimized for only light skinned models. As a result, dark skin colored people in photos were not visible due to over exposure. We need to think about the culture and values the data we select to train on shows.
Did I like this talk?
Yes, I enjoyed the examples provided and the organization of the talk. I may have already know the content but it was phrased in a organized and easily understood way.
Rook: Portable Storage on Kubernetes – Bassam Tabbara, Founder and CEO of Upbound, Inc.
- Kubernetes can be thought of as the cluster operating system or the container orchestrator. It also does life-cycle management, scaling, load balancing, scheduling, etc.
- It exposes a portable cloud API and thus supports the “write once, run everywhere”. You can for the most part containerize and deploy using Kubernetes and run in AWS, Azure, or other places fairly easily.
- When we start to manage storage beyond the lifecycle of a container (or other unit), we have to look outside of Kubernetes. Similarly, different forms of storage also need to be managed: blocks, volumes, queues, key-value stores, time-series, etc.
- Kubernetes allows portable volume extraction that is independent of pods. This is currently the only stateful workload abstraction.
- You need to interact with storage services that exist outside of your Kubernetes managed cluster for the most part.
- Rook is a cloud native storage orchestrator. It is an extension to Kubernetes hosted by the Cloud Native Foundation.
- It helps storage systems run like a service on top of a Kubernetes cluster by encapsulating storage system management logic.
- Can create a configuration that specifies size, monitors, and scaling that will create all components within the cluster that is needed using the operator pattern via Kubernetes operators.
- Rook operator defines a desired state and loops through state activating different components until the desired state is achieved and continues to do so to maintain desired state.
- This is an open source project looking for contributors!
Did I like the talk?
No. I understood what the problem and solution were but somehow I felt like I was missing something when he was explaining the examples. Most of them were “here is the configuration and then it just happens”.
Serverless and Kubernetes: The best of both worlds by Aparna Sinha, Google
- They key to a serverless platform is such that not many people are needed to maintain the servers and more time can be spent on developing the application.
- A lot of time in application development is spent on understanding monitoring, logging, orchestration, etc. that the application developer needs to spend time on.
- What is serverless at Google? Zero ops, auto-scaling, managed security; event-driven programming model, horizontally scaling with microservices; pay for usage.
- Serverless can be applied to more than compute including data storage and other components.
- Two main ideas: decouple applications and automate infrastructure
- This translates to: containerize the application + automate deployment + build in metering
- Kubernetes is container infrastructure as a service:
- It contains load balancing, networking, DNS and server discovery
- Environment agnostic management
- Has flexible, composable primitives that can be built together to craft your application. Ex. application + storage + deployments
- Kubernetes is a platform for automating management
- Can provide a desired number of replicas and Kubernetes will make this happen
- Provides auto-scaling based on system state like queue length or traffic rate
- Kubernetes is a declarative extensible API
- You can define different components or controllers you want to use in your own application
- Example: you could create a custom Spark controller that will allow you to treat Kubernetes as a Spark orchestrator
- Demo: the demo goes through creating an extension deploy a function via Kubernetes in such a way that makes the extension appear as part of the Kubernetes API. Using that function, write a hello.py function and deploy it via Kubernetes.
Did I like the talk?
Yes. I finally know what Kubernetes does (you’d think I would have looked it up by now). However, having worked with these types of orchestrators before it wasn’t too hard to figure out.
From Infra Ops to Platform Development: T-Mobile‘s Approach to Cloud – Nicholas Criss, Sr. Manager, Cloud Center of Excellence, T-Mobile
- Background: T-Mobile aims to be a market disrupting mobile carrier that strongly favors magenta the color
- Moving from waterfall development to agile development
- In 6 months moved up from last to first in JD Power survey
- Moved to the cloud and didn’t just follow “lift and shift” but tried to adopt the new philosophy of cloud development to make sure they didn’t fall behind competition
- Having already moved to the cloud, the company was seeing that they were able to adjust very quickly to sudden demands from business needs. Example: They were able to spin up tons of video processing servers very quickly and scale down as needed. Without being on the cloud it wouldn’t have been possible.
- Migrated through different stages:
- Before cloud: no cloud integration
- AWS console management: Manual management of cloud infrastructure
- Some automation of cloud infrastructure
- Managing infrastructure as code: all infrastructure was automated in code
- Software Developers: The software developers can easily (with minimal error) leverage fully automated infrastructure and eliminate the need for infrastructure teams
- Platforms: We got tired of having to relearn through repeating the same mistakes on different teams – make all these learnings a part of the platform. For example: scaling, fault tolerance, and general operational readiness.
- We moved from “we’re mostly security, compliance, and operations” to “we are software developers”
- Build our compliance and operations into code rather than Sharepoint docs that no one will ever read
- If you can’t implement something as code, it won’t happen or will happen incorrectly
- No more projects, only products – focus on building customer focused value instead of meeting deadlines and completing tasks
- Moved from being a telecom company that sometimes has interesting technology to a software company that happens to build telecommunications software
- Platforms that T-Mobile built internally to make software development better: Security management, policy enforcement platform, enhanced Hashicorp Vault service that optimizes for T-Mobile integration, enhanced container management platform, deployment platform
- Building out platforms to take away the painful parts of developing makes people more excited to develop and come to work
- They see these growth points as ways to interact with and give back to the community by open sourcing their projects
- Question: How did T-Mobile make this happen instead of just talk and no management support?
- From top level, management is supporting training and education
- From the bottom, there is the ability to innovate and remove pain by moving to a continuous improvement culture. There is a push to adopt new things to make old pain go away.
- There are still some holdovers of the old model with devops teams continuing to do deployments but there is progress and benefits being seen.
- Question: How do you get started by targeting small things to start building interest?
- Break off small pieces and build up the great infrastructure
- However, speaker encourages moving directly to going fully to containers and platform and don’t waste time through the in betweens
- Question: How do you migrate people to your platforms without the additional cost of making it work?
- We don’t force people to use a particular solution but instead we encourage them to use our solutions where they make sense
- It’s a continuous journey to make this work
Did I like the talk?
Yes. It was a rough start with some of the disjointed examples but it improved a lot by highlighting the journey steps and the growth that happened each time.
Diversity and Inclusion Panel Sponsored by Stripe
- Uma from Stripe Infrastructure Team manager
- Amy is a system software engineer at a local startup
- Dana is a manager at Google
- Q: How did you choose your field?
- Uma: I was one of the “accidental tech people”. In India, my grades were not high enough for medicine and the only other option was engineering. Software was chosen because it was less intense than the hard engineering. I found that I enjoyed it and continued.
- Amy: Tried to do pre-med and didn’t like it. Through a chat roulette interaction via Omegle, decided to try computer science!
- Dana: Father introduced her to computers and started playing with programming and got into that way.
- Q: What’s a problem in your field that you are excited about?
- Uma: Since we are growing to be global we are solving a lot of problems of scale and growth.
- Amy: Working on Kubernetes, we want to enable larger companies to be able to migrate easily to new technology.
- Dana: We focus on talking about how many 9s we want to support. We need to push harder to have better quality because with millions of users, even 0.01% is a lot of people.
- Q: Amy, what are some ways you can demonstrate leadership without being in a formal leadership role?
- Amy: I am not a manager and I don’t want to be one. I still want the ability to influence others. I achieve that through deep understanding and making my knowledge available through youtube or automating and sharing.
- Q: Dana, as you started moving into your tech lead role, what kinds of pros and cons were you thinking about? What advice was helpful?
- Dana: Tech lead can mean a lot of things and in silicon valley it can mean a tech manager. What I considered before moving into this position was to talk to other women in a tech lead manager position at Google. I recommend you make sure to talk to other women in these positions before making this move. What I got was: the hybrid tech manager role is a challenging one at Google. You won’t be able to do everything you want to do and you need to learn to let go of some things. However, you get a broad impact. You need to be comfortable not coding and still having an impact.
- Uma: You moved from being an engineer to manager and then from manager of engineers to manager of managers. What motivated you to do this switch?
- Uma: First, I didn’t plan to be a manager and was offered the opportunity to try being a manager after coming back from maternity leave. My manager supported me trying it out and would keep the IC option open if it didn’t work. I found that I really liked it and wanted to keep mentoring and growing other developers. Moving to the manager of managers position was due to seeking a new challenge with a level of abstraction. However, people don’t tell you how lonely it is being a manager since you lose the casual chat of other ICs.
- Q: When you were earlier on in your career was there support that you wished you had and how would you recommend people find it?
- Uma: Women need to ask more for access to things. Try to share what you learn with others to help them be aware of what it out there.
- Amy: The money that people are willing to offer you is a measure of how much they value you. Having a diverse group of people you can talk to to understand the potential range of values and opportunities that could exist. Try to find mentors or sponsors in different demographics and in different pay levels.
- Dana: I was lucky enough to get interesting products until I moved to Google. You need to find the thing you want to work on and make a difference on your own. Look for opportunities to do something you haven’t done before.
- Q: When it comes to negotiating, what advice do you have to figure out what to ask for?
- Dana: I was underpaid at my last job. Checked Glassdoor at a point where I was ready to potentially leave and asked them to pay me competitively or I would go. And they paid me more. Use recruiters as an opportunity to reassess you value and gain leverage.
- Uma: Continually research and update your data on what you should be paid. Don’t base your targets on your emotional response to the money. A lot of things are negotiable like working from home and you can negotiate for what is important for you.
- Amy: Always be on the job market to keep collecting data on what you are valued at. Make sure to try to get a pay band so that you are aware of how much you can negotiate for and where the lower end of tradeoffs might be. If you can, be on social media and get yourself out there.
- Q: Do you notice any behaviors that you wish under-represented groups would display more?
- Dana: If you haven’t noticed, I’m trans and had spent some of my career as a different gender. I have been socialized to have some of that other gender’s traits and as a result I don’t back down even though I have a feminine communication style. I will call people out and stand up for myself. I would like to see more women doing that. The way people behaved towards me when I changed to female. It was harder and I had to adapt. I encourage everyone to stand up for themselves.
- Amy: Ask to be treated equally and have the same as what your peers. Don’t change your communication style because other people don’t like it.
- Uma: Risk-taking in a more calculated way is something I’d like to see more people doing. Build a support network so that you can take risks.
- Audience Question: Do you think that offering training to help people learn these skills that would allow them to get an even footing culturally would help?
- Uma: Yes, at Stripe we had a training around negotiating that really helped understand that everything can be negotiable. Further, in backgrounds where saying no to managers is hard, building 1:1 relationships that helps understand the new ways interact with peers and manager.
- Amy: Managers should constantly be emphasizing that they are equals and not superior so that directs are comfortable say no or constructive discussion. It is really helpful that my managers communicates that I’m allowed to challenge them.
- Audience Question: How do you thank people for treating you like an equal?
- Dana: Thank them. It’s okay to tell them and explain to them why it matters to you.
Did I like this Panel?
Not particularly but that’s only because I’ve already heard this advice a number of times.
Fireside Panel with CIO of Providence Health Janice Newell and Charu Jain CIO at Alaska Airlines
- Q: Can you tell us about a service story that went horribly wrong?
- Charu: a lot of our cloud providers are on their own journey as well and even though you do everything right on your end, one day you find out that something isn’t there when you expect it to. We’ve learned to trust but verify even with some of the large providers?
- Q: How do you verify?
- Charu: Work with them to prove that it is working and pair on testing.
- Q: What about a great thing?
- Charu: We were building a cloud infrastructure to support more bookings. After moving to a cloud based solution we found that we were able to do things instantly that weren’t possible before.
- Janice: I have a horror story from us. During the Wannacry virus, we found out that one of our cloud providers that had advertised high security and a whole bunch of other features. It turns out they didn’t actually have those features. The virus then took down voice transcription for doctors for over a month. This created a huge problem for the customers. We ended up moving all of our physicians off of the SaaS based solution to an on-prem solution. We then went back to verify and inspect all the features and fault tolerance the company said they offered and were horrified to see how terrible it was. We pushed back and had to get them to improve their systems. This is an example of the trust but verify example is similar to what Alaska had to do.
- Q: We’ve seen a trend of a shift from Hybrid (on-prem + cloud) to an idea of multi-cloud where you use multiple cloud solutions (Azure + AWS). Are you thinking of multi-cloud and how would you use it?
- Charu: Our strategy is stated as multi-cloud. We are currently on Azure and have multiple SaaS providers. We are trying to get better at one thing before maturing into a multi-cloud environment.
- Janice: We are using containers to try to adopt that flexibility.
- Q: How has cloud adoption impacted things organizationally and culturally? Devops? Team makeup between specialist, generalist, or devops?
- Charu: We are working on a phased plan where we move one team and then take the learnings and roll that out to more teams. We are choosing teams first that have the most legacy technology that needs to be moved further. Developers feel empowered to drive these changes. We stood up automation teams to help start up the infrastructure and that has migrated to a cloud center where we help teams understand developing for the cloud.
- Q: Has there been resistance along the way?
- Charu: Change has always met resistance. People who have embraced it are setting the example.
- Q: What is your approach to monitoring?
- Charu: We are still in early stages. We are trying to make our monitoring complete but also not too noisy.
- We need to measure to know how our services are doing.
Did I like this panel?
No, it seems more managerial/business level than technical. There was a lot of repetitive back and forth so it was hard to pull out relevant information.
Machine Learning At Scale On Microsoft Azure, Paige Bailey, Microsoft
- Artificial Intelligence: any technique that allows computers to mimic human behavior
- This means anything that is coded to behave as a human which started off as rule based systems
- As the behavior becomes more complex and can’t be enumerated, you can use machine learning to let the computer learn a large space than the rules defined by a human
Types of Machine learning:
- Do you have labelled data?
- Yes – Cluster Analysis – Example: K-means
- No – Unsupervised Learning
- These problems are harder
- Do you want to Group Data?
- Example walkthrough: you are given a CSV with age, credit history, and employment; determine if you should give someone a loan
- You can make a decision tree to classify this person into yes or no buckets
- Aside: Jupyter notebooks are excellent for machine learning
- Feature: think of this as a column in a CSV
- Record: think of this as a row in a CSV
- Demo: showing a simple training example in a Jupyter notebook. (Author’s note: many tutorials are available online that will walk you through this)
- Machine learning is helping us learn without explicit rules; deep learning is now helping us identify features to learn on and extrapolate huge amounts of potential cases. Example: Magenta
- After this there were more examples and a discussion of how you can save your company money by learning to do machine learning on your own and also bump your salary due to the additional skills
- GPUs are very good for training neural networks
- Data science virtual machine using X2Go: if you don’t want to have to worry about managing your own machine learning environment, this will maintain all versions of libraries and languages to help you move faster with tons of educational Jupyter notebooks out of the box including topics like TensorFlow. They are ready, waiting, and free!
Did I like this talk?
Yes and no. I already knew all of this so it wasn’t super helpful but the presenter did a great job starting from 0 and showing people good examples in the time given.
Container and Serverless at Edge Connected to Cloud – Ying Xiong, Chief Architect, Cloud Platform, Huawei
- What is edge cloud? Computing resources at the edge managed by the cloud owned by provider or customers with bidirectional communication between both.
- Example scenario: security cameras where processing can be done on the camera itself and send data back to cloud
- Challenges: resource constraints and frequent network partitioning with no public access; we also need to support edge to edge communication
- Kubernetes on the Edge: what happens when and edge node loses the network connection? We can’t update it or support serverless on it.
- To solve these problems:
- We need to create robust communication between edge and cloud with high reliability and scalability
- Build an edge resource controller that manages the edge from cloud
- Support asynchronous metadata syncing and local storage to allow edge to operate on it’s own until connectivity is available
- Similarly, build a component called KubeEdge that will manage the edge applications and listen to events to allow self maintenance on edge nodes
- Use an event hub as a centralized listener to aggregate and queue events locally on the edge
- Created a custom TCP/IP to optimize for edge-cloud communication for reliable bi-directional multiplex channels that can scale
- Needs to support AuthN with certificates
- This provides a Kubernetes plugin to manage and maintain edge nodes and also provides Kubernetes on the node to support app status syncing
- KubeEdge is an optimized version of kubelet on edge for managing the edge node lifecycle and maintenance. It can work offline.
- Not yet open source but that is the plan.
Did I like this talk?
Yes. Before going to this conference I hadn’t hear the term “edge cloud” and it turns out it is very aligned with my technical interests.
Fireside Chat with Matthew Prince, Co-founder & CEO, Cloudflare
- When Cloudflare started we were competing with companies that were trying to find newer and better hardware to accomplish their goals; we did the opposite: we searched for ways of making the commodity hardware scale cheaply and linearly.
- Over 150 datacenters for Cloudflare with fully owned hardware in rented spaces; ISPs are interested in moving Cloudflare closer to them since it creates a faster internet and better customer experience.
- We have a world-wide infrastructure and we are suffering from low performance that isn’t improving from existing DNS providers due to lack of competition. We were able to create a secure and performant competitor in response.
- Discussion on opening business to public offering, costs and profits, competitive advantages, customer successes, various marketing and business data
- Insider threats are a big security challenge going forward
- “Our scale is so large that attackers can’t DDOS attack”
- Question: You made a decision to stop supporting a Neo-Nazi customer. How have you dealt with this internally?
- “I never thought I’d know so much about Neo-Nazis”
- “It’s lots of fun to kick Neo-Nazis off your service”
- There’s no way for consumers to understand when they are using CloudFlare. It seems creepy if we curate what users can or cannot see.
- We had to decide what the role of deep cloud infrastructure companies is
- The point at which we decided to do this is when the customer started to attribute their political view to the company
- There is still no clear answer on how to handle these cases
Did I like this talk?
No. This appeared to be a very long advertisement for Cloudflare. There were tidbits of interesting notes about network structure. Some really great quotes came out of this one.
Power Talk by Peter DeSantis Vice President, AWS Global Infrastructure and Customer Support followed by a Fireside Chat
- At a large scale you are able to invest in things that wouldn’t be cost efficient at lower scale
- AWS has expanded to many region and at an accelerating rate
- What is a region? A set of availability zones?
- What is an AZ (availability zone)? A fully isolated infrastructure with one or more datacenters, a unique power center, and a meaningful distance from each other
- Every region has at least 3 and up to 6 AZs
- Between datacenters within the AZ
- Between AZs within the region
- Between AZs and the public internet across the region
- Built on top of a dense fibre backbone with high redundancy than Amazon manages
- Also creating inter-region Direct Connect that allows customer to get a fast lane between regions
- Compute with AWS EC2: Security, Performance, and Familiarity (look and feel is like any other server)
- Going over the history of EC2 architecture
- First started off with customer instances on top of a hypervisor that managed security, network, and hardware
- In the next big iteration, they moved over to Nitro infrastructure
- Through these moves they were able to improve network latency by 20%
- This also enabled moving to commodity hardware – except for security and some other core features
- Could continue to look for hardware solution, use FPGA chips, or go with a custom ASIC – eventually acquired Annapurna labs to realize the last option
- Thus equipped, security improved and allowed nearly 100% of compute resources to be available to customers on the instance. This achieved the familiarity goal.
- Now have also opened offerings for bare metal instances as well as Nitro VMs
Did I like this talk?
Yes. Despite knowing a lot of this information at a high level, I hadn’t seen in presented in this particular nutshell. I liked how it was phrased and the progression through the topics.
- The event used a phone app to support planning your agenda, quick connect with other attendees, integration with sli.do for live polls
- Regular video on projectors but also a live keyword note taking screen which was actually very helpful
- The breakfast was great: hot toasted bagels and a hot oatmeal bar for breakfast
- Event support for developers is great: they provide desks and tables with a high density of plugs available
- The sponsors presented before, between, and after each talk. It was a little overwhelming and felt like I was watching Youtube with ads turned on. However, the comfort of the event was worth it. Similarly, all “ads” were relevant so it wasn’t too bad.
- Very easy to network at this event: all I needed to do was sit down at a table and people would start up conversations about technical topics
- Despite trying to avoid doing so, people are very interested in hearing the terrible things about the company I currently work for
- I encountered some discussion because I perhaps foolishly brought my motorcycle gear with me (I did ride the motorcycle to the event) and started some conversations about whether or not to lock your helmet to your bike. This then lead to discussions about companies that let you work from home and some nice connections. If you learn one thing from this: bring or wear an item that will inspire people to talk to you.