2023 - Year In Review

As a software engineer, I helped build a few systems and products at work this year. Here I present a high level overview of my learnings and thoughts without specificities to the products I built.

Published On: 19 Dec 2023

TL; DR ->

1. Webhooks

If you are building a notification-based system where you want to notify any external entity about an event that happens in your system, webhooks are the way to go. Webhooks are simple HTTP endpoints (mostly POST) that are registered at the site of the sender. The sender makes HTTP calls to the endpoint whenever an event happens. The sender can implement authentication mechanisms and retry logic as well. I was at the receiving end of the webhooks; our job was to consume the data that was sent to us by a third party.


Since events can happen at any time, it is important for the receiver to be up always. The sender can implement retry mechanisms in case the receiver is down (mostly indicated by 502 GatewayTimeout, but can be any other error as well), however, it is good to catch the events the first time. Some senders can also stop sending events when they get too many errors. So, we had our receiver deployed in Kubernetes (GKE) with scaling and health check configs defined. And, we have never observed a downtime ever since.

No throw new InternalServerError()!

500 InternalServerError is the scariest. If the server does not have good error logging (which is not the case most of the time), debugging them is very hard. But imagine yourself throwing 500s! We made that mistake initially using 500 as a catch-all error code. Eventually, webhook sender started poking us about 500s they observed. We eventually had to replace all 500 throws with appropriate error codes with the following rules in mind:

  • If the error is from the sender side, return a non-success status, such as 400 BadRequest. The sender can retry the message, which hopefully will become successful
  • If the error is from our side (the receiver) and we can’t do anything to fix the error, there is no point in asking the sender to send the message back. Thus, send a success status and log the error so that someone/some process can review and take further action on it

Handling PII

PII is Personally Identifiable Information such as the name, email and address of an individual, around which there are strict laws about how to handle them. The sender was sending us PII, which we did not require and did not want to be captured in any of our systems, including logs. Fortunately, we had turned request logging off in the Ingress, and with the help of Zod, were able to filter only the fields that we required.


Zod takes care of defining the schema and validating whether the data follows the schema. While the sender in our case has defined a schema, it was unclear and quite messy. We had to do trial and error in production (yes, we missed some data) to get the schema correct and sometimes broaden it. To give out a concrete example, there was an ID field which was sometimes a number and other times a string 😅


Since the webhook receiver is a simple public HTTP endpoint, anyone can send data to it. This is a huge problem and can cause data pollution in our case. Webhook senders can rely on several mechanisms to authorize their message (i.e., we can be sure that it is them who sent us this message)

  • HTTP Basic Authentication (yeah, it says authentication, but can be used for authorization as well) - Sent as base64 encoded username:password string in the Authorization header (see, now it is authorization again - HTTP Standards 😁). Sender and Reciever agree on username and password which are kept secret. These credentials are sent along with every request and do not change with each request.
  • HMAC Authorization: HMAC stands for Hashed Message Authentication Code. The basic idea is simple - 1. take the message and a shared secret, 2. concatenate them, 3. take the hash of it and send it along with the message. The receiver with the knowledge of the message and the shared secret can repeat the same computation to derive their own hash. If the hashes are equal, the authenticity of the message is established. The actual algorithm goes one step further to make the operation more robust. The important thing to note here is that the credential (the hash) changes with every message.

In both of the cases, it is important to keep the secrets secret. An adversary with the knowledge of secrets can forge a message without the knowledge of the receiver. Cryptography can protect us, but we need to protect the keys 🔑. In our case, the sender implemented HMAC authentication and we got the secrets from them. We stored the secrets securely and made sure that it is never logged and sent in the API responses of our system.

Beware of Middlewares

When we implemented HMAC, the local tests were all working fine, but all messages were rejected with 401 Unauthorized in production. We were confused as to why this was the case, and the culprit turned out to be the body parser. When our Fastify server gets a message, it parses the body based on the Content-Type header. In our case, it was application/json and we got an object back in the response handler. We converted it to JSON before calculating HMAC. But it turned out that we needed access to the raw body. We hooked into the appropriate place and were able to add a rawBody: Buffer | string to the request object. This body was sent to HMAC and we were able to get the correct hashes.

Message Processing

We deploy webhooks to listen to messages sent by a third party. But what is the point of setting them up without handling the data we receive? Where to do the data processing is an important architectural decision to make. If the data can be processed fast enough, the webhook route can contain the processing logic. Oftentimes, this is not the case. When the processing time increases, the sender ends up waiting for the response and might potentially experience timeout issues, causing them to retry more. This is a bad situation to be in, as the new message requires further processing. We ended up putting the messages in an MQ and sending a successful response immediately. The MQ consumer can handle the messages at its own pace without impacting the webhook. This also enables us to scale the webhook and the message processor independently.

Getting rid of groupBy aggregations

The concept of groupBy and realtime data do not fit in nicely. We do not know whether we are finished receiving a group yet or not when we are operating in a real-time mode. The groups sometimes need to be stored in memory, which might cause performance bottlenecks. We have been experiencing this problem for a long time, and moving to real-time data processing using webhooks motivated us to solve this problem once and for all.

We were fortunate that our database model allowed us to generate the aggregate incrementally, rather than needing to commit everything at once. We ended up doing row := f(row, newChunk) instead of row := g(chunk1, chunk2, ...), where f and g represent the operations on the data, by compromising a few assumptions about our system. Again, engineering is about choosing a perfect trade-off; we got rid of waiting and huge memory usage by switching to real-time mode by sacrificing a little assumption and more DB writes.

2. Implementing a Distributed Cache!

One of our NodeJS applications was doing heavy DB reads on the tables which did not change often. We implemented a simple in-memory cache helper that can be used something like const cachedFunction = cached(function) with all the type-safety. The cached function also introduced a function invalidate on the cachedFunction, which can be used to empty the cache.

As our use cases grew, we used the cached utility to wrap the controllers backing our UI data. Everything was running fine in production until we decided to increase the number of replicas of our application in k8s. UI users started reporting inconsistent states in the UI, where the resource they edited was not being successfully shown back in the UI (yeah, we did a GET request to refresh the UI after a POST to save, we were lazy to build lazy UIs for this internal application)

We immediately got to know that the issue was with the cache, and decided to move it out of the memory and share it across the replicas. There was an in-memory/external hybrid cache built in Scala by the veterans that relied on message passing to bust out in-memory cache entities while updating the external cache simultaneously. Since we did not have the luxury of using Scala code in our NodeJS application (Scala to WASM? 🧐), we decided to go with a simple route of using only external cache backed by Redis.

We initially went ahead with Redis Dictionary with the key cache/${functionId}. JS provides Function.toString() to get the code of the function (look, no fancy Reflection 😛), which we hashed to get the functionId. Then, we took the JSON of the argument, and stored it as a dictionary key, with the JSON of the result stored in the value. While it worked fine, it suffered with respect to invalidation. Redis does not support expiring individual keys of the dictionary to keep life simple. So it was an all-or-nothing approach - expire the entire function cache or keep all of them, which surely is not flexible.

In the end, we resorted to using plain Redis key space with keys being in the format cache/${functionId}/arg1/arg2/... where argN is the JSON representation of the function argument. Our cache looks like a tree now with the branches occurring whenever the arguments change. With this, we can expire individual keys. To do so, we parametrized the invalidate function to take a prefix of function arguments. For example, if we have a function f(a,b,c), if we call invalidate(a), it will prune the Cache Tree for all arguments starting with a. If we call invalidate(a,b,c), we target a specific element. This tree structure allowed us to retrospect on function prototype design - while you can have getRows(tableName, columns, filter) and getRows(filter, columns, tableName) doing the same thing, the order matters if this function is cached - we want the invalidations to be minimal (If we can have 3 args in an object, the concern is gone and the code becomes more readable too!). As an added bonus, the invalidate function is also type-safe - In f(a,b,c), if a was of type string, you can’t pass number as the first argument to invalidate (expect an article on how we did this 🙂). While this implementation is solid and flexible, everything comes with a tradeoff and we ended up doing an O(n) SCAN + DELETE during invalidation.

3. Going Local-First (Use your Apple Silicon!)

Developer Productivity is important for any company. CI/CD exists to test and deploy a commit without the developer needing to do anything, but having a preview while development always helps. Most, if not all, modern web frameworks come up with a way to run a local server with a live preview. This might be challenging to build in a work environment due to dependencies on external entities such as databases with a copy of production-like data. While running an entire database on a local machine is possible by restoring the production backup, it takes time and is hard to keep it up-to-date. Here is a solution that I came up with at my workplace:

  • Running low coupled dependencies using docker-compose - We run Redis, Memcached, RabbitMQ and sometimes Postgres as well
  • Tunnelling some dependencies using ssh - assuming the dependencies are accessible in a VM, and you can access the VM remoteVm, it is easy to tunnel the connection to your local machine - ssh -N -L localPort:remoteHost:remotePort localPort2:remoteHost2:remotePort2 remoteVm (Thanks to a colleague for pointing this out!). We forwarded Postgres most of the time, along with the HashiCorp Consul for remote service discovery.

This was an easy setup and we wrapped it in a NodeJS script (we could’ve done it in bash, but I couldn’t give up await Promise.all() to run processes in parallel). This script is invoked in the dev script of every project, making life easy. We were able to save hundreds of developer hours with this setup, thanks to containers and tunnels.

4. Hi, Monorepos

At some point in our time, our NodeJS project grew big and we had several utilities in it. New NodeJS projects were on the horizon where these utilities could be obviously useful. Here we hit a branching in the path, we could either

  • Publish utils as a NodeJS package to a package registry and use it in other packages, or
  • Have utils as a package in a monorepo and use it in other packages in the same monorepo.

Both of the options have pros and cons - as any other options in tech do. Publishing a package needs its own CI, and comes with versioning nightmares if not done correctly. But they allow multiple users to use different versions of the library (Not sure if this is a Pro or Con – in my opinion, it is definitely a Con and creates tech debt). On the other hand, a monorepo avoids all the configuration needed apart from setting it up and forces all the clients to use the latest version of the libraries. This makes library authors mindful of not introducing any breaking API changes and new behaviours unless absolutely needed. Read monorepo.tools for good coverage on this topic. We ended up choosing turborepo and pnpm workspaces for us. We needed to slightly modify our container builds to not include any unnecessary dependencies. All in all, we now have 3 libraries and 4 projects in our monorepo, with the number of the projects expected to grow further.

5. Building NodeJS Build Tooling

We embrace Type-Safety for JavaScript (we also do care a lot about not typing the things explicitly where it can be inferred - who wants Car car = new Car()?) – so TypeScript is our go-to choice. That comes with an additional step of compiling TypeScript to JavaScript, where tsc gets us covered. Over time, when moving to monorepos, we figured out that we needed libraries with multiple entry points (allowing us to import {dbClient} from "library/database", import {mqClient} from "library/mq", for example). Plain tsc gets us a one-to-one mapping from ts to js files in the dist folder, with us needing to edit the exports field of package.json manually. What if we could "exports": "dist/*.js"?!

Welcome to the world of bundling, where it is common in front-end builds. Take a React application - if not bundle split, all your logic lives inside index.js, which is loaded by index.html with just a script tag pointing to the js file. We chose ESBuild for bundling our project, with an entryPoints: string[] defining the list of the files that we wanted to expose. ESbuild takes care of bundling their requirements together, compiling TypeScript to JavaScript (it does not do any type-check, we need to run tsc with noEmit before we invoke the actual build). As an added bonus, we also got a watch capability for the dev servers. We haven’t tackled watching the workspace dependency libraries yet, since it has not been a problem for us till now.

For the frontend bundling and dev server, we use Vite and could not ask for a better solution. We follow an opinionated folder structure reflecting the UI routes. We use this information to split our bundle into multiple chunks per a top-level path, which can be loaded in parallel by the browser when loading index.html.

6. Building a Shopify App

Shopify is a platform for anyone to build an e-commerce store with management capabilities and nice UX out of the box, which can be further personalized according to the merchant’s requirements. Shopify allows to extend its functionality via third-party apps, which can integrate into lots of surfaces within it, ranging from admin-facing UI to user-facing UI. The apps need to be hosted somewhere but can choose to render on Shopify’s Admin UI inside an iframe (called as Embedded Apps) or externally (called as Non Embedded Apps). I got the opportunity to build a Non-Embedded App this year.

Getting to build things from scratch where the requirements and deadlines are clear but nothing else is is interesting. You get to be the UI/UX designer and a Product Manager thinking of the user and the flow, You get to be the infra person defining the CI/CD processes, and ultimately you are the Full Stack person building the application itself – yeah, jack of all trades. The core challenge for an engineer building an external integration is the understanding of the target APIs, their requirements and the constraints. Shopify shines in this aspect here by providing excellent developer docs at shopify.dev. Their GraphQL playgrounds are also excellent and help test things out before actually coding them. So, it was an absolutely seamless experience.

The Tech Stack

A Shopify External app is simply a web app that uses Shopify’s APIs to communicate with a Shop to read and modify its state. As such, we could have gone with a Full-Stack React framework like Remix, to avoid moving back and forth between API routes, API handler in the UI, and the UI component itself. Due to some constraints, we decided to go with the exact architecture we wanted to avoid - Fastify NodeJS server, Vite React Frontend sitting inside a monorepo (compromises are inevitable in tech, but we did not end up in a severely bad position, which is okay). For new projects, Shopify’s template comes with Remix.

Shopify APIs

Shopify provides GraphQL and REST APIs, with the former being preferred. The version their APIs with yyyy-mm format and have clear deprecation policies. One common source of frustration is that the response schema is vastly different between GraphQL and REST APIs for the same concern. Their NodeJS client SDK also provides useful wrappers around the most commonly used APIs, such as auth and billing.

Type Safe GraphQL

Both TypeScript and GraphQL embrace the concept of type safety, but it is sad that they do not embrace each other due to their differing philosophies (Hi tRPC!) To solve this, we need to have some sort of typegen that can help TypeScript understand the types. We ended up downloading Shopify’s GraphQL Schema definition, hosting it in our repo (it is a big fat JSON, sad to the ones who count code lines as metrics) and using @appolo/client to do the typegen. They provide a gql function to begin with (this is a normal function, not a tagged-template literal function with the same name - it was one of the confusing points) and the typegen looks for the named queries and mutations being passed into it to generate the types (being named is an important point here). This makes the call sites type-safe by providing type information for the variables and return types. We made type generation run along with the dev server and the build script, along with making generated types not committed to the main repo.

Handling Rate Limiting (Spoiler: We didn’t)

Shopify’s Admin GraphQL requests are rate limited via a leaky bucket mechanism. We get pre-allocated tokens for every time window and each query and mutation consumes some of them. If a query needs more tokens than we have, we get a rate-limiting response. We had use-cases for pulling large amounts of data using Shopify APIs, but building a system that is robust and correct is interesting and challenging, and it meant us tackling engineering problem instead of a business one.

Fortunately, we managed to bypass the rate-limiting by using Shopify’s Bulk APIs, which execute a given query and notifies us with the export of the data once the execution is complete. On some other occasion, we managed to look up the data from within our own database instead of calling Shopify APIs.

I think the moral of the story is that being lazy (read: thinking out of the box) helps to expand the horizon of possible solutions, which might be more plausible and cost-effective sometimes. We could have built a perfect rate-limiting solution, but it could have taken a few weeks in development and testing and would have required some maintenance due to an edge case no one had imagined.

Implicit State Machine

Every software and hardware is an implicit state machine. Since our app involved a flow-based onboarding mechanism, it made more sense for us to store them in the DB as {stepADone: boolean, stepBDone: boolean, stepCDone: boolen} (no, there were no booleans apart from some. I am oversimplifying the things there - I hear you screaming that I could’ve used an enum here!). Then, an API call gets this state in the React Router loader. Based on the current state we are in, it redirects the user to a certain page. We embraced the React philosophy of UI = f(state), from the component level to the application level.

We made sure that a user can not land in an invalid state by making sure that:

  • executing routing logic at the top level loader - if a user enters a URL corresponding to a not-yet-accessible state, the loader will just redirect them back to the state they are in.
  • not to include <Link/>s for these states anywhere.

After all the measures in place, we could not solve the back-button problem, where the browser takes the user to the previous page and skips executing the loader, we made a compromise - it was a feature for the user to edit whatever they have done in the previous step, not a bug 😅

Resource Creation and Idempotency

At one point in the flow, we had to create a resource in the backend which should be unique. The API to create that resource had no such restrictions and happily created another resource with a new key. Fortunately, we got to understand this problem during development, thanks to React.StrictMode calling useEffect twice (you might yell at me saying I might not need an effect, but believe me, we had to implement that in a page with only a huge spinner in it. We triggered the React router action as soon as the user entered the page using the effect). We came up with the following solution:

  • Enter a mutex and execute the following operation
    • If we have the key in the DB, we do not proceed to create and just return the key
    • If we don’t have the key, call the API and store its key in the DB and return it

The Mutex is the key here preventing multiple external API calls from happening. We can use Postgres’ pg_advisory_xact_lock or Redis RedLock to implement a distributed Mutex.

Overall Result

The project is live at Shopify Marketplace with a few dozen happy customers - apps.shopify.com/truefit (This is the only place I brag about my workplace in this article). The project was an accelerated idea to prototype to production. While the acceleration provided us with a lot of ideas and opportunities, it also left a few blind spots only to be uncovered by the users using it. We have since maintained a log of the issues and a disaster management & recovery plan. After a few months, the project is in a stable landscape now.

7. Murphy’s Law

I got introduced to Murphy’s Law by a colleague, of which one version states as:

Anything that can go wrong will go wrong, and at the worst possible time

This is especially true in Software Engineering and UI design where we make a lot of assumptions about the system that we are building and the users using it. Users’ instinct might be vastly different from the Engineer’s one - that’s where the UI/UX team comes in to fill in the gap. When it comes to a backend system, we might get hit with an unexpected downtime that breaks our system and assumptions. Over time, I’ve learnt to handle such situations by:

  1. Not making too many assumptions about the system, but ensuring we do not over-engineer anything. We can’t live without assumptions about the world that we live in
  2. Handling the errors reactively by patching the system, while ensuring the patch does not introduce additional assumptions or break the system in any way

We have experienced lots of such situations reinforcing our disaster management strategies - who wants to be woken up from sleep just to execute one command to fix things? - just build CI/CD and add docs!

8. Cooking a Chrome Extension with Bun and React

Bun was released this year with the promise of being performant and providing solutions to commonly used things out of the box. At the same time, we got a requirement to collect the data through an injected JavaScript from some websites. Ideally, the websites include our script which collects data (like analytics data, but not analytics data 😜) and sends it to our server. In development, we had to inject the script ourselves into the webpage – Enter Browser Extensions.

Browser Extensions is a standard that defines a set of APIs to allow programmatically enhance the behaviour of the browser using plugins. We ended up choosing Chrome Extensions, which provides a global chrome object to access these APIs inside the extension. An Extension is a regular JavaScript code running inside the browser in an isolated context along with HTML and CSS for the Sidebar and Popup UIs. We wanted to build the extension with TypeScript, though - fortunately, there was @types/chrome package for us. Apart from the code, we need to define a manifest.json which contains meta-information about the extension, along with the API versions it would like to use - Remember Chrome deprecating Manifest V2? there was a webRequest API which allowed extensions to intercept and modify network requests and responses. In Manifest V3, it is not there anymore and is replaced by declarativeNetRequest, where the rules are declarative and do not allow an extension to peek into the request (privacy-preserving…Google…)

scripting API

We used chrome.scripting.executeScript to execute our code and get the result back in the Popup surface. However, soon, it was painful for us to write document.getElementById("result")!.innerHTML = JSON.stringify(result) where we used to write it declaratively in React. Fortunately, Bun eliminates the need for build tools and bundlers - we can use tsx files to write React code and compile them into JavaScript using bun build (We need to install react and react-dom though). Then, all the things were easy - we ran executeScript inside an useEffect, and stored the result in a state, which is used to render the result. Thanks to React, we can bring fancy JSON Viewers into the mix.

declarativeNetRequest API

Moving on, we started injecting our script through our existing integration code patched by a proxy server. This needed us to redirect the request that was hitting prod servers to our dev machine. We started patching /etc/host to point the prod domain to the local one, and soon ran into HTTPS issues. Even with the self-signed certificate, Chrome refused to allow the connection to localhost 😅. A quick Google search yielded us Requestly, which solved the problem of redirection for us. We went a step further and integrated the redirection into our extension using the declarativeNetRequest API. We can write the request editing config using a JSON specifying the request matching criteria and a handler action. We used RegEx filters and Substitutions to achieve our end goal. We further added functionality in the extension UI to turn the rules on and off with the help of a toggle (Read: Radio button - we had 3 options - local, staging, none).

Bun Impressions

Bun made things smooth for us, although we ended up using esbuild in a js file for the build and dev server steps. A major impression for us was the package installation time, it is rocket speed when compared to npm and even pnpm. Apart from that, their standard APIs are also thoughtfully built, and, compatibility with node is one of their major selling points (which deno did a lot later to catch up with). We even wrote CI steps using bun, reducing the installation time and speeding up the CI. Fast feedback times are always important for a developer and thanks Bun for helping achieve that!

9. Using the Web Standards

Developing in JavaScript touches both the backend and frontend worlds. While the backend has many runtimes, frontend’s web runtime is now fairly standardized. During the course of this year, I used several Web Standard APIs to achieve few tasks at work. Here is a summary of them:


The fetch API is fairly standard nowadays, allowing us to easily communicate with the servers without the hassle of managing the XMLHTTPRequest state. However, fetch calls are cancelled when the user navigates away from the page. While this is not a problem for normal API calls needed to populate the UI, it presents a problem for tracking events where we could potentially lose some of them.

navigator.sendBeacon() provides an alternative to POST API calls sending the tracking events. The body can be plain text or a Blob with a content type. Browser ensures that the navigator calls are finished even after the user leaves the page, making it perfect for sending analytics events. It is logged as a ping event in the Browser’s Network Tab and can be inspected as any other request.

I was sending application/json created in a Blob through sendBeacon and observed it was making a preflight OPTIONS call to get the CORS headers. Switching body type to text/plain would not do so. CORS became a real issue for me during local development with proxies set by declarativeNetRequest, so we ended up using plain text and a server-side parser z.string().transform(bodyString => JSON.parse(bodyString)).transform(bodyObject => schema.parse(bodyObject)). A nice thing about fastify is that any exception thrown at any point in this chain would be returned as a 400 BadRequest.

✨ While I was writing this article, I learnt that fetch with keepalive: true would achieve the same effect as navigator.sendBeacon()! It would allow us to specify mode: "no-cors" to send the data and not to be aware of the response, which is fine in some cases.


One of our tasks during the data collection script is to observe changes to the DOM, and MutationObserver is the go-to solution. With Mutation Observer, we can observe a DOM node and get updates via a callback whenever it changes. We can observe the changes in attributes, text data and children; optionally for the subtree as well.

We started observing a few nodes with predefined selectors for the changes on them. However, we lost the event when the node itself was removed from the DOM. Thus, we decided to observe the document and execute our handler if a mutated node matches() our selector. We carefully decide whether to execute the callback or not through a series of short-circuit conditions to exit as early as possible.

Vary header and Caches

When I was writing a proxy to patch the existing script to add our injection on the fly, I got access to a Readable stream from the proxy which I can read, patch and send back to the client. However, the requests made from the browser returned gibberish on the stream, leaving me puzzled. For a moment, I wondered if I was reading a binary HTTP response. We hit a lightbulb moment when we made the same request from curl and we got a plaintext stream! Turns out the server was returning different responses based on the Accept-Encoding header, which we removed in the proxy to always get plaintext responses back.

This behaviour where the response can change based on a request header can cause caching issues if not configured correctly. Say a client can only accept plaintext and we return them the gzip which they can’t process. This situation is avoided by using the Vary header on the response which caches respect (both CDN and Browser), which can be set with a list of the request headers based on which response varies. In our example, Cache can store different entities for plaintext and gzip response if Vary: Accept-Encoding.

Referer based policies

When we want to customize the script by the website it loaded from, we can follow a few strategies:

  • Generate a query parameter unique for each website and ask them to include it in the script src. However, this does not prevent another party from using the same query parameter if they know its value 😅
  • Derive the information about where the script was loaded from the request itself - The Referer header comes into play here

While we can inspect Referer header from the request and customize the response based on it, some user agents can skip sending this header due to privacy concerns. We ended up using hybrid of both the approaches to achieve required customization.

Writing some CSPs

Hosting a website/application in public can come up with its own security challenges. What prevents anyone from embedding your website in a full-screen iframe, hosting it in a similar domain as yours, and having overlays to steal credentials? The answer is CSPs (Content Security Policies), which we can set in the Content-Security-Policy HTTP response header.

For the case that I mentioned above, setting frame-ancestors 'none' does the job, but CSP headers can achieve much more than that, for example, it can restrict where the images, scripts and iframes in the website can come from. Choosing an ideal CSP depends on a particular setup, and is crucial.

10. Advanced Data Structures

Solving a problem at work and plugging in a solution from your Data Structures coursework is one of the most satisfying moments. We had a lot of requests containing the same data but wanted to process them only once. While the Set is a perfect solution for this problem where we process a request if we don’t see it in the Set yet, they can be prohibitive in terms of memory at our scale. What if we can trade some accuracy off for way less memory?

Enter Bloom Filters - they are a probabilistic set data structure, which hash the incoming element into a bitset. This means the elements can’t be retrieved once we put them back into a Bloom Filter - an ideal solution for PII such as IP address. We can ask the Bloom Filter the question of whether we have seen a particular element or not, and its response has the following characteristics:

  • When it says no, we can be 100% sure that we have not seen the element yet
  • When it says yes, it can be wrong with a probability p, which can be pre-configured

It made sense for our use-case to lose a few elements at the cost of seeing them the next day, as we rotate bloom filters every day. We used Redis implementation of Bloom Filters, which provides handy commands to interact with it. To store 100M elements, we only consume around 200MB with an error rate of 0.0001, which is perfectly fine for us.

11. Year of AI - Following the Hype?

Hi ChatGPT, you changed a lot of behaviours when it comes to information retrieval from the internet. While I was initially sceptical to adapt ChatGPT at work, GPT-4 Turbo is doing absolute wonders with added capabilities of browsing and code execution. In my experience, ChatGPT can automate trivial and boring tasks - I asked ChatGPT to browse declarativeNetRequest docs and come up with the JSON rules satisfying our use-case, although I had to tweak a little bit later. I follow a rule where I never submit any PII to the ChatGPT - not even the name of my workplace.

Choosing the Right tool for the Right Problem

We also had a few interesting problems to solve with the AI, one of them was identifying variant size options from an apparel website. We started with Question Answer models by feeding them the HTML content of the website and asking them questions about size options - they were slow and inaccurate, and might’ve needed fine-tuning - we did not pursue that way further. ChatGPT and other LLMs were good at this, but sometimes hallucinated and were slow and costly - it was like using an axe to kill a fly. We turned out to simple embedding models as our last resort.

Data Matters

Turned out we had a huge dataset of such tokens at our behest, albeit not clean. We ran our tokenizer and normalizer which removed special characters, converted everything to lowercase and split the words on this data. We also deduplicated these tokens ending up with several thousands of them. By seeing the dataset now, a few tokens did not make any sense based on our domain knowledge, which originated from garbage data. We ended up joining a meeting, going through these tokens and manually deleting the tokens which did not make sense, sometimes double-checking to understand where they originated from. At the end of this 1-hour-ish exercise, we had gold clean data!

FastText - So Fast, So Huge!

FastText is an embedding model by Meta’s FAIR Lab, and lives true to its name. It generated word embeddings for our entire dataset within a matter of seconds on the CPU. We initially trained FastText on our dataset and observed a lot of embeddings having 0 norm or 1 norm - it was a problem with our dataset where the tokens did not occur frequently as we deduplicated it during the generation phase. We also observed that their pre-trained models had a good understanding of our domain data, which we verified further by taking the embeddings and visualizing them in the embedding projector. We were happy to see related tokens forming clusters of all shapes and sizes - a dream of a data scientist. At that point, we decided that the pre-trained model was good for our use case. The only sad thing was that the pre-trained model weighed around 6GB and we have yet to figure out how to fit it inside a docker container and make it scale.

VectorDB and Inference Pipeline

Now that we have decided on the model, we took the embeddings of our dataset and stored it in an in-memory vector DB using pynndescent to do an approximate nearest neighbour search (no fancy VectorDB yet, sorry folks!). Then, for every input token, we pass it through the model, take the embedding and check if we have a nearest neighbour to it in our dataset within a predefined distance. If we find that neighbour, we accept that the token is valid, otherwise reject it.

We hosted a flask server to execute this logic and made our JS call the endpoints defined in the server. The JS filters the HTML by some common-sense assumptions about where the size variants are located in a website. This drastically reduces the number of tokens that we need to process. We empirically found that this approach gets the result that we want most of the time, but we have yet to come up with benchmark methods and the numbers - a task for the coming year!

LLMs are not the Silver Bullet

An important learning from this model-building exercise was that LLMs can not solve all the problems, we can be better off with small, domain-specific models that are trained on our data. LLMs are costly to execute and can be very slow, but they also hallucinate - which can be a huge problem depending on the use case.

But I’m also excited about RAG (Retrieval Augmented Generation) in the form of Function Calling in OpenAI API, followed by Gemini. We can hook up the right tools to the LLMs to do specialized tasks and present their result in the form of natural language - this presents huge opportunities for interfacing them with the real-world, though some caution is always helpful.

12. Ops Mode

Being a FullStack Engineer means I’ve to touch the Infrastructure handling as well to get things going. We use GCP at our workplace and we have a bittersweet experience with it as everyone has with other cloud providers – no cloud provider is perfect. Following are some of my experience and thoughts on a few areas of GCP, their k8s offering GKE and general Infra provisioning.

Where are the modules, MemoryStore?

MemoryStore is GCP’s Redis offering, with promised autoscaling and SLAs. Who wants to manage things when there is a managed solution? We use MemoryStore for most of our Redis instances. Unfortunately, MemoryStore does not support Redis Modules and we can’t live without our Bloom filter solution. While we can use a library that implements Bloom filters on top of the Redis bitset, their module looked lucrative to us and we ended up deploying a helm chart on our GKE cluster to move ahead with the MVP. We have to figure out monitoring and scaling parts. I was impressed by the PersistentVolumeClaim handled by GKE seamlessly.

BigQuery - Big Query

BigQuery is GCP’s Data Warehouse solution where the storage and computing are horizontally scaled across their clusters. While horizontal scaling offers faster query computation times, it can cost money if the queries are careless about the data they are accessing. We limit the compute scaling by using a predefined number of slots, which helps to keep the costs predictable.

The only query interface to the BigQuery is SQL Query in Google SQL syntax. While SQL is declarative, I observed queries tend to grow big especially if we have to do data deduplication and joins. The query lives true to its name - Big Query.

One of the nice features of BigQuery is that we can route the application logs to a dataset with minimal configuration with the help of Log Sink. The Log Sink creates its own dataset and tables with the schema to match the log schema and takes care of synchronizing the tables with the logs. We use this in all places - just log to stdout using a structured logger such as pino, and rest assured that this data ends in BigQuery. I would imagine using fluent-bit and its transport to achieve a similar solution for more customizability.

Often, we need to process this log data, aggregate it and write it to a destination table. While Scheduled Queries can help us achieve this functionality with a minimum delay of 15 minutes, they do not come up with a retry mechanism. We observed scheduled queries fail due to a concurrent transaction modifying the same table. This is a huge problem for us and we are planning to move to Jenkins for more control on when queries are executed and retry mechanisms. Google, do something for it!

Ingress Provision Time

When shipping a product to provision, making it available to the end-users is the most exciting thing. When you have all the available things deployed and the endpoint takes forever to be up, it adds to the frustration. Add more time, we started to worry if any of the configurations had gone wrong. Add more time, everything is up and running. I’ve observed some Ingress provisions taking up to 30 mins. What?! - Shall we contain the pressure to make a big announcement?!

I understand a lot of things happen when provisioning an Ingress. A VM has to boot up and configure itself, and all the routers need to be configured across the request chain etc., But with my previous experience hosting ALBs in AWS through EKS Ingress Controller, GCP’s one is less delightful. Can Google do anything to address it and provide more real-time updates?

HCL, you are too restrictive!

We use Terraform to provision and manage our Infrastructure. While HCL (HashiCorp Configuration Language) is expressive enough, it is not expressive as a programming language. It is okay for a Configuration Language to be more restrictive than a programming language, but the requirements can be anything. Can you think of filtering JSON keys by value? I need to go for esoteric-looking comprehensions in HCL, which can be very much expressed fluently in a programming language such as TypeScript. While HCL functions cater to most of the things, they are limited when compared to a rich ecosystem of libraries such as npm.

This is a difference between philosophies of course - what a Configuration Language is supposed to be, whether it should be Turning-Complete or not etc. Personally I tend towards Pulumi, due to following reasons:

  • Enables to write a clear and concise code following the language’s best practices
  • Enables to adapt the language ecosystem of the language to do hard computation - such as diving the IP into subnetworks and calculating the masks
  • Enables to adapt the standards of the language w.r.t modules and code-reusability. I can think of an infra monorepo with pnpm
  • Enables to provision a set of resources (stack) dynamically based on an API call! This is a real killer if your use case demands this functionality

All in all, it is important to understand the chain events to lead to the current state in a workplace, and adapt to it. Of course, HCL is not bad and is loved by many people!

Hi FluxCD - Let’s do GitOps 🚀

We had a custom-script-based deployment system at our workspace, which traditionally deployed jars to VMs and managed the process lifecycle. We naturally extended the scripts to support k8s deployment in the same style as our legacy deployments, but, found them too much restrictive as k8s can naturally handle lifecycle better than us. With that, the search for a k8s-native deployment solution began.

We started with FluxCD and ArgoCD, both of which are in CNCF and are great. Since we preferred the CLI style of deployment and the state more tightly tied to a git repository, we chose FluxCD. We now have a flux configuration repository managed by Terraform and a deployment repository containing configs and kustomization overlays for all our applications across the environments. The CODEOWNERS file ensures that no one is privileged to deploy to prod without code approvals while giving enough freedom to deploy in lower environments.

We now have to create a branch, do a commit, merge and forget way of deploying our software with Flux taking care of everything else. We had a Flux issue one time - it was not able to pull in the deployment repo changes, which was easily fixed. We plan to smoothen the process more by providing CI hooks to automate deployments (We use GitLab CI) and flux diff views of MRs which can ensure reviews that the MR is not doing unintended.

Concluding Remarks

This was a summary of what I did in work this year. This is not exclusive and omits lots of small details — our embrace of the functional programming world in JS, handling fine-grained reactivity using signals in React, going through thoughtful architecture designs and discussions etc. I want to stress that none of this is possible without the great team, management and product - and, I thank them all. Connect with me on LinkedIn to understand more about the work we do, and how we help people (This was brag-ish, not completely though!).

This is my first time writing a summary of the year. This writing process helped me to reflect on this year and celebrate the learnings and achievements - I think this is important to keep you motivated to keep on doing great things year-long.


I discuss what I did in work this year, with 12 sections, summarized as:

  1. Building webhook receivers, and the best practices regarding response handling, data processing, auth and PII handling
  2. Implementing a distributed cache for NodeJS servers, and the emergence of the Cache Tree and its pruning (I won’t leave any spoilers for this one. Go read about it!)
  3. How to make best use of your work machine by running software locally, and how I helped implement that
  4. Solving package distribution and out of date package tech debt using Monorepos
  5. NodeJS bundling for the back-end code and our journey with esbuild
  6. Building a Shopify app in an accelerated manner, and the learnings along the way
  7. Murphy’s law and how we learned to prepare for and handle failures
  8. Building a Chrome Extension with Bun and React, and exploration of some of its APIs (we built a Proxy!)
  9. Few Web Standards we utilized along the way
  10. When the Data Structure course learnings manifest at work - Bloom Filters
  11. AI AI AI!
  12. Usual Ops Rants - Read them!