Charles Xu
Essays, books, wiki on technologies, career, markets, and more.
Category git
We all know Google employs a version tracking system that uses a single repository/depot. Every close-source google product that you love is tracked by this single repo, which is so...
Git is one of the most widely used version control systems. Traditionally, a repository on git is considered as a complete history of the entire project in the form of...
Committing often and pushing often has been advocated as good practice when using Git, which saves your latest work on remote even if your hard drive dies right after and...
Category web
Pods in Kubernetes are the smallest orchestration unit and are ephemeral by definition: Deployment/StatefulSet/DaemonSet/ReplicaSet updates or patches Nodepool downscaling (compaction) or upgrades (cordoned and drained)
Summary To achieve zero-downtime service update, Kubernetes rolling update implies the API must be both forward and backward compatible. Forward compatibility is hard if at all makes sense. Blue-green update...
Imagine you run a single page app at example.com that communicates with backends over restful API and is authenticated with JWT tokens managed by you, but identities are managed by...
Imagine your app is doing tremendously well with growing traffics. If there is a single server for your app, and the server is approaching its capacity, how would you scale...
Message queues are an asynchronous inter-process communication protocol that gains much of its glory with the recent hypes in microservices. Senders and receivers do not interact with the middleware at...
HTTP provides an abstraction of short connections. Unlike the continuous byte streams in TCP, exchanges between client and server over HTTP starts with a client request and ends with server...
A common design for content display, pagination partitions information into multiple pages and serves one at a time. We have seen it in search results, message history, and cascading news...
I have been working with another three wonderful people on the senior design project, which is a web application of an inventory management system, and the production is up at...
All teams will henceforth expose their data and functionality through service interfaces. There will be no other form of inter-process communication (IPC) allowed: no direct linking, no direct reads of...
Category microservices
In a large, busy cluster, how do you scale Istio to address Istio-proxy Container being OOM-Killed and Istiod crashes if too many connected istio-proxies?
Imagine you run a single page app at example.com that communicates with backends over restful API and is authenticated with JWT tokens managed by you, but identities are managed by...
As presented in my previous post, microservices are the state-of-the-art architecture for building scalable, highly-available, manageable backend. No more 30-minute build time, single point of failure, and constant regression from...
Message queues are an asynchronous inter-process communication protocol that gains much of its glory with the recent hypes in microservices. Senders and receivers do not interact with the middleware at...
HTTP provides an abstraction of short connections. Unlike the continuous byte streams in TCP, exchanges between client and server over HTTP starts with a client request and ends with server...
A common design for content display, pagination partitions information into multiple pages and serves one at a time. We have seen it in search results, message history, and cascading news...
I have been working with another three wonderful people on the senior design project, which is a web application of an inventory management system, and the production is up at...
All teams will henceforth expose their data and functionality through service interfaces. There will be no other form of inter-process communication (IPC) allowed: no direct linking, no direct reads of...
Category distributed systems
A super quick recap, one’s complement represents negative x by reverting every bit of x, while two’s complement negative x as one’s complement of x plus 1. Symbolically,
Imagine your app is doing tremendously well with growing traffics. If there is a single server for your app, and the server is approaching its capacity, how would you scale...
In the world of internet, sending messages in clear text is like swimming naked. We would love some secure communication channels free from eavesdropping or tampering. Security as such is...
Git is one of the most widely used version control systems. Traditionally, a repository on git is considered as a complete history of the entire project in the form of...
Here is where we stand. Fisher-Lynch-Patterson has shown that consensus is not guaranteed in bounded time in a purely asynchronous network. The CAP theorem shows that from consistency, availability, and...
Remote Procedure Call (RPC) is a canonical structuring paradigm for client-server/request-response services.
Category signal processing
We all share the sorrow and misery from that signal processing class. You were thrown at some crazy formula, kind of know how to use them but probably never understand why...
Category networking
In a large, busy cluster, how do you scale Istio to address Istio-proxy Container being OOM-Killed and Istiod crashes if too many connected istio-proxies?
AWS EKS on VPC networks need AWS Security Group Rules (SG) to receipt ingress traffic. But what if you reach the max rules count in your SG?
Your Kubernetes cluster probably has a shared ingress for north-south traffic, coming from a cloud load balancer and lands on your favorite proxies like Envoy, or Istio gateways, or Nginx....
We go from containers and network namespace to Pod-to-Pod, Pod-to-Service, and external-client-to-Service networking.
Kubernetes does not differentiate sidecars and application containers in a Pod. Hence, enabling Istio for short-running workloads imposes additional challenges to the conventional approach of injecting an Envoy sidecar to...
DNS prefers UDP. There are times when DNS must run on TCP (request or response size exceeds a single packet, perhaps due to too many response records), but UDP is...
My team at Cruise operates tens of Kubernetes clusters with 10,000s cores and 100s of TB of RAM. Since migration to GCP, we have hit several interesting scaling issues. One...
A super quick recap, one’s complement represents negative x by reverting every bit of x, while two’s complement negative x as one’s complement of x plus 1. Symbolically,
Imagine you run a single page app at example.com that communicates with backends over restful API and is authenticated with JWT tokens managed by you, but identities are managed by...
In the world of internet, sending messages in clear text is like swimming naked. We would love some secure communication channels free from eavesdropping or tampering. Security as such is...
Here is where we stand. Fisher-Lynch-Patterson has shown that consensus is not guaranteed in bounded time in a purely asynchronous network. The CAP theorem shows that from consistency, availability, and...
Category istio
In a large, busy cluster, how do you scale Istio to address Istio-proxy Container being OOM-Killed and Istiod crashes if too many connected istio-proxies?
Kubernetes does not differentiate sidecars and application containers in a Pod. Hence, enabling Istio for short-running workloads imposes additional challenges to the conventional approach of injecting an Envoy sidecar to...
As presented in my previous post, microservices are the state-of-the-art architecture for building scalable, highly-available, manageable backend. No more 30-minute build time, single point of failure, and constant regression from...
We all know Google employs a version tracking system that uses a single repository/depot. Every close-source google product that you love is tracked by this single repo, which is so...
Category security
In the world of internet, sending messages in clear text is like swimming naked. We would love some secure communication channels free from eavesdropping or tampering. Security as such is...
Category docker
Introduced in version v17.05, multi-stage builds feature in Dockerfiles enables you to create smaller container images with better caching and smaller security footprint. Fundamentally, the new syntax allows one to...
Many regard containers as a virtualization technology. They are missing out. Docker has much more to offer. It is a graceful solution to some of the most painful experience in...
Category kubernetes
In a large, busy cluster, how do you scale Istio to address Istio-proxy Container being OOM-Killed and Istiod crashes if too many connected istio-proxies?
AWS EKS on VPC networks need AWS Security Group Rules (SG) to receipt ingress traffic. But what if you reach the max rules count in your SG?
Your Kubernetes cluster probably has a shared ingress for north-south traffic, coming from a cloud load balancer and lands on your favorite proxies like Envoy, or Istio gateways, or Nginx....
We go from containers and network namespace to Pod-to-Pod, Pod-to-Service, and external-client-to-Service networking.
My team at Cruise operates tens of Kubernetes clusters with 10,000s cores and 100s of TB of RAM. Since migration to GCP, we have hit several interesting scaling issues. One...
Pods in Kubernetes are the smallest orchestration unit and are ephemeral by definition: Deployment/StatefulSet/DaemonSet/ReplicaSet updates or patches Nodepool downscaling (compaction) or upgrades (cordoned and drained)
Summary To achieve zero-downtime service update, Kubernetes rolling update implies the API must be both forward and backward compatible. Forward compatibility is hard if at all makes sense. Blue-green update...
It is common to extract the application configuration to a separate file as a runtime dependency of the container image that includes the application binary. As a result, the same...
Introduced in version v17.05, multi-stage builds feature in Dockerfiles enables you to create smaller container images with better caching and smaller security footprint. Fundamentally, the new syntax allows one to...
Category operation
I hope you find inspirations from these pretty neat shell tricks and my shell setup.
Pods in Kubernetes are the smallest orchestration unit and are ephemeral by definition: Deployment/StatefulSet/DaemonSet/ReplicaSet updates or patches Nodepool downscaling (compaction) or upgrades (cordoned and drained)
Summary To achieve zero-downtime service update, Kubernetes rolling update implies the API must be both forward and backward compatible. Forward compatibility is hard if at all makes sense. Blue-green update...
It is common to extract the application configuration to a separate file as a runtime dependency of the container image that includes the application binary. As a result, the same...
Category career
Make sure to check out the previous post: Advices I wish I got at the start of my career.
Disclaimer: Opinions are my own. Not investment advice.
This post explains the expectation of each engineering level in the most concise and company-agnostic way and reveals the steps towards promotion.
Unlike in school when we get grades on every assignment and in every course, we get less frequent feedback in professional life, usually once or twice per year, which is...
When I was a kid playing chess with my dad, he sometimes would offer me hints on some good moves. I would never make those moves. I would rather make...
Category go
Unlike C++ or Python, Go does not support function parameters with default values if unspecified. Specifically, we want that
Category cloud
In a large, busy cluster, how do you scale Istio to address Istio-proxy Container being OOM-Killed and Istiod crashes if too many connected istio-proxies?
AWS EKS on VPC networks need AWS Security Group Rules (SG) to receipt ingress traffic. But what if you reach the max rules count in your SG?
Your Kubernetes cluster probably has a shared ingress for north-south traffic, coming from a cloud load balancer and lands on your favorite proxies like Envoy, or Istio gateways, or Nginx....
My team at Cruise operates tens of Kubernetes clusters with 10,000s cores and 100s of TB of RAM. Since migration to GCP, we have hit several interesting scaling issues. One...
Category investment
Before Fundraise: Allow minimum three to six months to raise money. Have a clean cut from last job to avoid IP disputes. Prepare data site (Certificate of Incorporation, Bylaws, board...
Disclaimer: Opinions are my own. Not investment advice.
Category startup
Careful planning and execution work for general management but not for startups. Perfect execution is futile if you end up building something nobody wants (waste). The real progress for startups...
Before Fundraise: Allow minimum three to six months to raise money. Have a clean cut from last job to avoid IP disputes. Prepare data site (Certificate of Incorporation, Bylaws, board...
How to do product-led growth and hands-on outbound sales at the same time?
Streamlit, about to raise its Series-C, was acquired by Snowflake for $800M in March 2022. In this conversation with Adrien, we chatted about OSS metrics, licenses, open-core vs freemium vs...
Notes derived from a guest lecture by Danny Wallace, Partner at PwC’s Silicon Valley practice. For informational purposes only. Errors and omissions are my own.
Notes on Intellectual Property (IP) law for founders and busy professionals. Not legal advice. For informational purposes only. Laws can change, so this article may contain dated information. Always consult...
Category oss
Streamlit, about to raise its Series-C, was acquired by Snowflake for $800M in March 2022. In this conversation with Adrien, we chatted about OSS metrics, licenses, open-core vs freemium vs...
Category artificial intelligence
Many inference speedup techniques mirror the classic systems regime—such as caching, paging, tiling, pipelining, and speculative execution (e.g. branch prediction and cache prefetch). Speculative decoding, generalizing speculative execution to stochastic...
In the multi-head attention mechanism, why after reshaping the projection matrices for Q/K/V from 3 dimensions to 4, we need to transpose the tokens dimension with the heads dimension?
Category llm
Many inference speedup techniques mirror the classic systems regime—such as caching, paging, tiling, pipelining, and speculative execution (e.g. branch prediction and cache prefetch). Speculative decoding, generalizing speculative execution to stochastic...
In the multi-head attention mechanism, why after reshaping the projection matrices for Q/K/V from 3 dimensions to 4, we need to transpose the tokens dimension with the heads dimension?
Never miss a story from me, subscribe to my newsletter