Computer scientist building the infra AI needs to scale around search engines, databases, and distributed systems.
Earth
Joined August 2012
- Tweets11.9K
- Following389
- Followers783
- Likes63.6K
Which of these will you use to share cached build outputs between your laptop and CI?
→ Turborepo
→ Nx
→ Bazel
→ Something else
1
7
Bazel can cache the output files from individual build steps for reuse across machines: bazel.build/remote/caching
2
Before your coding agent installs any packages, your machine may execute an attacker's code during a repository clone.
GitHub describes this recursive-clone flaw in its 10 November advisory. Update Git to 2.56.1 or a patched release for your series.
2
15
Update applications that bundle Git too, because they may use a separate copy.
github.blog/open-source/git/…
4
Customers may need an open browser tab to receive the work that your AI feature produces.
For work that customers can collect later, use a background job that saves its output so that the app can retrieve it. Inngest, for example, lets the app retrieve a job's status and output after the work finishes.
Before you deploy, request a report in the test app and close the tab during generation. After the job finishes, reopen the app and confirm that it shows the report without requesting another generation.
1
8
Inngest documents how the app can retrieve job status and output with an event ID.
inngest.com/docs/examples/fe…
2
Which of these will you use to count visits to your app's website without using cookies?
→ Plausible
→ Fathom
→ Umami
→ Something else
1
15
Umami is an open-source tool that counts visits to your website: umami.is/
2
Your database backups may fail while customers can use the app.
In its 10 November report on October's breach, Railway describes a backup failure in which services use outdated storage credentials while customers have access to their apps.
When you replace a storage credential:
→ Update every service that uses it, including the backup service.
→ Test the replacement credential before you deploy the new settings.
→ After deployment, confirm that the backup service creates a fresh backup with the replacement credential.
Sirannon's backup destinations guide shows how to store a backup and verify the recovered copy: sirannon.sondelali.com/docs/…
3
31
YOU CAN WIN A PAYMENT DISPUTE AND PAY 30% OF IT FOR THE HELP.
Stripe lists its Smart Disputes fee for AI-prepared responses as '30% of the disputed amount for each dispute you win.'
Compare that share of each recovered payment with the fee for a manual response and the time that you'll spend preparing it.
Will you pay a share of each recovered payment to automate disputes, or prepare the evidence yourself?
ALT A diagram divides an illustrative $100 recovered payment into a $30 Smart Disputes fee and $70 before other fees. The diagram states that the example applies to a dispute that you win.
2
11
Stripe also charges the dispute received fee, which applies whether you win or lose.
stripe.com/pricing
6
If your AI agent exceeds the model's context limit after a search, check how much text the search tool returns.
Ask your coding agent to measure the text that each search adds to the next model request, using the longest documents that your customers search.
This guide shows how you can fit the passages into a budget without cutting out the answer: narsil.sondelali.com/docs/ag…
2
20
Does your landing page let visitors use the app with sample data, or does every demo require an account?
ALT A dark square card shows the sentence: Give visitors a working demo with sample data before you ask them to sign up. The words 'a working demo with sample data' appear in blue.
2
2
40
Which base image will you use for a Node.js production container that needs native libraries?
→ Node.js Debian slim
→ Node.js Alpine
→ Google Distroless
1
1
71
Google's Distroless images contain the runtime and its dependencies, with no shell or package manager in the standard images: github.com/GoogleContainerTo…
13
Did you know that your coding agent can delete an AWS database that its connector's custom rules forbid it to delete?
AWS reports this bypass in its local AWS API MCP server, so update that connector to version 1.5.1 or later.
9
8
222
AWS's advisory covers local servers with custom policies; its hosted service and the blanket read-only setting are unaffected: github.com/awslabs/mcp/secur…
19
Customers can cancel an AI reply while your server keeps requesting the rest of it at your expense.
The Vercel AI SDK's Stop button cancels the browser's request, while cancellation on the server needs separate configuration. For replies that customers won't resume later, make the chat function forward the cancellation signal to the model request.
Before you deploy:
→ On Vercel, use the Node.js runtime and enable supportsCancellation for the chat function.
→ Protect required database saves with waitUntil or after so that the function can finish them after the customer cancels.
→ Cancel a long reply in a preview build and check that the app cancels its request to the model provider too.
1
116
The AI SDK guide explains both the server connection and the extra setting that Vercel requires: ai-sdk.dev/docs/advanced/sto…
7
When you have to record and edit your app's product walkthrough yourself, which of these will you use?
→ Loom
→ Screen Studio
→ OBS Studio
→ Something else
2
25
Screen Studio records your screen on a Mac and adds automatic zooms to the recording: screen.studio/
9