Python Bytes cover art

Python Bytes

Python Bytes

Written by: Michael Kennedy and Calvin Hendryx-Parker
Listen for free

Python Bytes is a weekly podcast hosted by Michael Kennedy and Calvin Hendryx-Parker. The show is a short discussion on the headlines and noteworthy news in the Python, developer, and data science space.Copyright 2016-2026 Politics & Government
Episodes
  • #489 Or JSON?

    #489 Or JSON?
    Jul 21 2026
    Topics covered in this episode: django-orjsonBest Django Redis configuration for speed and sizeLinus Torvalds puts the foot down against Anti-AI Kernel MaintainersDjango Steering Council backs the Triptych ProjectExtrasJokeWatch on YouTube About the show Sponsored by us! Support our work through: Our courses at Talk PythonConsulting from Six Feet Up Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedInCalvin: Mastodon / BlueSky / X / LinkedInShow: Mastodon / BlueSky / X Join us on YouTube at pythonbytes.fm/live to be part of the audience. Usually Tuesday at 7am PT. Older video versions available there too. Michael #1: django-orjson Adam Johnson dropped django-orjson - drop-in replacements for the Django and DRF pieces that touch JSON, swapping stdlib json for orjson, the Rust-based library. Headline numbers: 10x faster serialization, 2x faster deserialization.The interesting question is why this needs to be a package at all. pip install orjson is the easy part. Adam's actual pitch: adopting it "isn't easy, especially when your framework uses json in many different parts." Django scatters JSON across JsonResponse, the test client and test case classes, the json_script template tag, and more. There's no single hook to grab, so you get a library that catches them all.Adam is refreshingly honest about the scale of the win. His words: "While database queries tend to dominate the typical Django application's runtime, the time spent in serialization and deserialization can still be significant." He calls it "a nearly free performance win" - not "this will 10x your app." That's a claim about cost, not magnitude, and it's worth keeping those straight.Worth flagging what the post doesn't cover: caveats. There are none in the article, but orjson has real ones. Django and Flask both render datetimes as RFC 822 HTTP-date (Wed, 15 Jul 2026 12:00:00 GMT); orjson does ISO 8601. It can't do ensure_ascii, it rejects NaN and Infinity (which stdlib happily emits), and it raises on Decimal. If you've got a JS client parsing dates, that's a wire-format change.Who should actually take this? If you're a DRF shop shoveling JSON all day, yes - it's cheap and it's real. If your app mostly renders HTML templates, you're optimizing a slice of runtime that's already near zero.The problem Adam's package solves doesn't exist in Flask or Quart. They already centralize every JSON operation - jsonify, request.get_json(), the test client, the |tojson filter - behind one provider object at app.json. So there's no library to install. It's about ten lines: import orjson from quart.json.provider import JSONProvider # or flask.json.provider class OrjsonProvider(JSONProvider): def dumps(self, obj, **kwargs) -> str: return orjson.dumps(obj).decode() # provider must return str def loads(self, s, **kwargs): return orjson.loads(s) app.json = OrjsonProvider(app) The numbers on talkpython.fm Evaluated it, measured it, and skipped it. The biggest JSON payload we serve is our MCP server returning a cached episode transcript, about 139 KB. Swapping the provider saves 0.119 milliseconds per request. That total response takes 1.1 msWe got 4.1x, not 10x - and the reason is the good lesson. Payload shape decides your speedup. The 10x is for structure-heavy data, lots of small keys where stdlib burns time in Python-level dispatch per item. Our hot payload is one giant transcript string, so the work is escaping and memcpy Calvin #2: Best Django Redis configuration for speed and size Peter Bengtsson revisits a classic: his 2017 "Fastest Redis configuration for Django" benchmark now has a 2026 update posted this week.The 2017 post pitted django-redis serializers (json, ujson, msgpack, pickle) and compressors (zlib, lzma) against each other; conclusion was msgpack + zlib as the sweet spot - avoid the json serializer, it's fat and slow.The 2026 update narrows focus to just compressors: default (no compression), zlib, lzma, and newcomer zstd.New results: lzma compresses best but is slowest; zstd is the fastest compressor on Ubuntu; differences between them are very small.Big takeaway across both: compression buys you a lot of space (2–3.5x smaller) for very little speed cost - worth it for Redis where memory is the constraint.Caveat from the author: results depend heavily on your data - his test stores short strings of numbers, so benchmark your own workload. Michael #3: Linus Torvalds puts the foot down against Anti-AI Kernel Maintainers Write up on Ars.Really good coverage by Maximillian: Time to wake up (for some)Torvalds said that “Linux is not one of those anti-AI projects, and if somebody has issues with that, they can do the open-source thing and fork it. Or just walk away.”I agree with Max, putting your head in the sand and waiting for AI to go away will likely mean you won’t be working professionally in software development in the coming years.The statement came amid a lengthy thread arguing about the use of Sashiko, an “agentic ...
    Show more Show less
    31 mins
  • #488 tau - it's 2pi and it writes code

    #488 tau - it's 2pi and it writes code
    Jul 14 2026
    Topics covered in this episode: The trusted-publishing debate: how to do it right vs. why you shouldn't trust itJupyterLab 4.6 and Notebook 7.6 are out!Tau – new small, readable terminal coding agentDjango Tasks and Django 6.1ExtrasJokeWatch on YouTube About the show Sponsored by us! Support our work through: Our courses at Talk PythonConsulting from Six Feet Up Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedInCalvin: Mastodon / BlueSky / X / LinkedInShow: Mastodon / BlueSky / X Join us on YouTube at pythonbytes.fm/live to be part of the audience. Usually Tuesday at 7am PT. Older video versions available there too. Finally, if you want an artisanal, hand-crafted digest of every week of the show notes in email form? Add your name and email to our friends of the show list, we'll never share it. Calvin #1: The trusted-publishing debate: how to do it right vs. why you shouldn't trust it https://snarky.ca/how-to-publish-to-pypi-using-github-actions-securely/ (Brett Cannon) and https://blog.yossarian.net/2026/07/07/You-shouldnt-trust-trusted-publishing (William Woodruff) Trusted Publishing (PyPI's OIDC-based auth scheme, also now used by npm, RubyGems, crates.io, NuGet) replaces long-lived API tokens with short-lived, auto-scoped credentials tied to CI/CD machine identity.Yossarian's post: it's purely an authentication mechanism between a machine identity and a package — it says nothing about package safety or quality. PyPI deliberately avoids any "verified/trusted" badge for it, unlike its verified-URL checkmarks.Same logic applies to PyPI attestations: anyone can sign with any machine identity they control, so an attestation's presence isn't itself a trust signal.Bottom line from that post: don't confuse "trusted" (machine-to-machine) with "trustworthy" (human judgment about the package).Snarky.ca's companion piece is more practical: given GitHub Actions compromises in the news, the real fix is 3 concrete steps — run zizmor to lock down workflow permissions/checkout credentials and pin actions to commit hashes, adopt Trusted Publishing to eliminate stored PyPI tokens, and require manual approval via a GitHub environment before any publish job runs.Takeaway for listeners: Trusted Publishing is good hygiene for how you authenticate to PyPI, but it's not a substitute for securing your CI pipeline itself — or for actually vetting the packages you install. Michael #2: JupyterLab 4.6 and Notebook 7.6 are out! Michał Krassowski's rundown - a chunky minor release: 68 features, 97 bug fixes, 95 contributors, one of the biggest ever. Scratchpad console (Notebook 7.6 headliner) - a console next to your notebook sharing its kernel, for throwaway experiments. Ctrl+B.Jump to last-edited cell - new commands hop through recently edited cells.File browser glow-up - Date Created column, editable breadcrumbs with Tab-completion, and Open in Terminal.Debugger - sources open in the main area, floating step/continue overlay, live kernel-sources filter.Custom layouts (Lab) - activity bar top/bottom, draggable panels, four-way tab splits, per-panel Ctrl+scroll zoom.~5x faster extension builds - webpack → Rspack, and jupyter-builder means no full Lab install needed to build extensions.Keyboard/a11y - add shortcuts from the UI (no JSON), Find & Replace in Edit menu (Ctrl+H). Calvin #3: Tau – new small, readable terminal coding agent Tau – new small, readable terminal coding agent (Python 3.12+), built as both a working tool and a teaching project for how coding agents work under the hoodInstall via uv tool install tau-ai, pipx, or pip; ships a tau CLIThree-layer architecture: tau_ai (provider-neutral model layer) → tau_agent (reusable "brain": messages, tools, events, loop) → tau_coding (CLI/TUI, file & shell tools, sessions)Supports OpenAI, Anthropic, OpenAI Codex, OpenRouter, Hugging Face, and custom/local OpenAI-compatible endpointsBuilt-in tools (read/write/edit/bash), durable JSONL sessions with resume/branching, project instructions via AGENTS.md, and context compactionCore harness is UI-agnostic — same brain can power the TUI, print mode, or a custom frontend — usable as a standalone library too Michael #4: Django Tasks and Django 6.1 Django 6.0 finally ships first-party background tasks (django.tasks) - out of Jake Howard's DEP 14, accepted May 2024, after two decades of everyone bolting on Celery/RQ/Huey.It's an API, not a worker. Django handles task definition, validation, queuing, and result storage - it does not execute them. You bring the backend.The default backend traps people. ImmediateBackend runs tasks inline on the request thread and blocks until done - so out of the box .enqueue() backgrounds nothing (a 5-second task means a 5-second response). The other built-in, DummyBackend, runs nothing at all. Both are dev/test only.Nice API otherwise: slap @task on a function, call .enqueue(), get back a TaskResult you look up later by id - with async twins like aenqueue(). Gotcha: ...
    Show more Show less
    32 mins
  • #487 Minimum requirements

    #487 Minimum requirements
    Jul 7 2026
    Topics covered in this episode: dust - a better duHermes Agent: The AI agent that grows with youllm-coding-agent 0.1a0ExtrasJokeWatch on YouTube About the show Sponsored by us! Support our work through: Our courses at Talk PythonConsulting from Six Feet Up Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedInCalvin: Mastodon / BlueSky / X / LinkedInShow: Mastodon / BlueSky / X Join us on YouTube at pythonbytes.fm/live to be part of the audience. Usually Tuesday at 7am PT. Older video versions available there too. Finally, if you want an artisanal, hand-crafted digest of every week of the show notes in email form? Add your name and email to our friends of the show list, we'll never share it. Michael #1: dust - a better du du + Rust = dust - a fast, visual, intuitive disk-usage CLIRun dust and immediately see the biggest directories and files without piping through sort, head, or awkSmart recursive output focuses on what matters instead of dumping every folderColored bars show relative size and parent/child hierarchy, making “where did the space go?” obviousPerfect for Python projects bloated by .venv, caches, Docker volumes, downloaded datasets, and local AI modelsInstall via brew, cargo install du-dust, conda-forge, Scoop, Snap, deb-get, or GitHub releases Calvin #2: A Way better ARchive format for Python packaging war - new archive format spec from Astral (same team as uv/ruff), v0.0.2, still no binary encoding defined yetHeader-Index-Store layout: header IDs the file, index maps names to store offsets, store holds compressed dataIndex uses a finite-state transducer (FST) to dedupe common path prefixes across entry namesSupports three entry types (file, directory, link) and three compression modes (store/DEFLATE/zstd), plus an "executable" metadata flagUnpacking is atomic - writes to a temp dir, then renames into place, so a failed extract never leaves a half-unpacked directoryStrict name-segment rules (no NUL/control chars, no leading/trailing whitespace, blocks Windows-reserved names like CON/PRN) to avoid path traversal and cross-platform footguns Michael #3: Hermes Agent: The AI agent that grows with you Hermes Agent is an open-source, Python-built AI agent framework from Nous Research - think ChatGPT-style assistant, but connected to your tools, files, shell, browser, calendar, memory, and messaging appsI’m using it in Discord as a long-running agent conversation, not just a one-off chatbot sessionHermes can connect through a gateway to platforms like Discord, Telegram, Slack, WhatsApp, email, webhooks, and more - so the same assistant can follow you across surfacesIn my setup, I can send Hermes voice/text from Discord, keep project context across turns as threads, and ask it to actually do things: read GitHub repos, run commands, edit files, schedule calendar events, generate drafts, and verify resultsA fun workflow: I can trigger one-shot actions from an Apple Watch shortcut - dictate a request, send it to Hermes, and have the agent execute it asynchronouslyHermes has persistent memory, so it can remember durable preferences and facts - for example, how I like my research formattedIt also has “skills,” which are reusable procedures the agent can load later, so Hermes can self-improve over time instead of rediscovering the same workflow repeatedlyIt supports scheduled jobs / cron-style automations, so it can proactively watch for releases, send summaries, run checks, or remind you about thingsIt’s provider-agnostic: OpenRouter, Anthropic, Google, xAI, local models, Nous Portal, and othersThe big idea: Hermes turns an LLM from “a chat box I visit” into “an agent I can reach from anywhere that knows my workflows and can take real actions and learns over time.” Calvin #4: llm-coding-agent 0.1a0 Simon Willison built a Claude/Codex-style coding agent on top of his llm library, using an alpha of the llm package plus his python-lib-template-repoBuilt almost entirely via prompted TDD - asked an agent to write a spec.md, then commit + implement with red/green tests, occasionally hitting a real OpenAI key to sanity-checkShipped to PyPI as an alpha: uvx --prerelease=allow --with llm-coding-agent llm codeTool set mirrors familiar coding-agent primitives: read_file, edit_file (exact string replace + diff), write_file, list_files, search_files, execute_commandAlso exposes a Python API - CodingAgent(model="gpt-5.5", root=..., approve=True).run(...) - which Simon didn't ask for but got anywayDemo: llm code --yolo told GPT-5.5 to build a SwiftUI CLI clock; model correctly noted SwiftUI isn't really CLI-friendly and still produced an ASCII-art time display Extras Calvin: Slides, but for developers https://sli.dev/Wanna reduce your token usage…. only issue is that its lossy https://github.com/teamchong/pxpipePEP 772 - Python Packaging Council inaugural election dates set, nominations open July 28, voting September 1-15 Michael: What the pls? revisited! Joke: Min requirements for ...
    Show more Show less
    28 mins
adbl_web_anon_alc_button_suppression_t1
No reviews yet