Python Bytes

Michael Kennedy and Calvin Hendryx-Parker

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.

  1. hace 4 h

    #498: A Tiny Episode

    Topics covered in this episode: MemTensor / MemoryOS PyPI package hijacked via a malicious build backend TinyMongo Jev: what to know One innocent dict read makes attribute access permanently slower Extras Joke Watch on YouTube About the show Sponsored by us! Support our work through: Our courses at Talk Python Consulting from Six Feet Up Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedIn Calvin: Mastodon / BlueSky / X / LinkedIn Show: 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: MemTensor / MemoryOS PyPI package hijacked via a malicious build backend On Sept 23 an attacker published backdoored MemoryOS 2.0.34 on PyPI and three bad versions (0.1.21, 0.1.23, 0.1.25) of MemTensor's OpenClaw plugin on npm. PyPI had no clean release that day, so 2.0.34 was the newest. They pushed commits to MemTensor's own GitHub Actions release pipelines. On PyPI that was a custom Poetry build backend, and on npm a tweaked validation script. Both used BASH_ENV to hand the publish token to the attacker before the real publish ran. SafeDep couldn't confirm how the attacker got push access. Runs on import, not install: A Go implant called sckit starts when the library loads, so --ignore-scripts won't save you. It harvests credentials from your home directory (npm and PyPI tokens, GitHub tokens, SSH keys, cloud CLI tokens, .env files) and sends them to skyleen[.]fr servers. It's a worm: It uses stolen tokens to copy itself into other repos and packages, so the victim list could grow. If you installed it: Downgrade to MemoryOS 2.0.33 (plugin 0.1.20) and rotate every credential reachable from $HOME. Also kill any running sckit stage0 process and check repos you can push to for a stray runtime-update.yml workflow or .sckit/ directory. Michael #2: TinyMongo Want to use a MongoDB data interface, but swap out the storage engine? Memory for testing/caching JSON/TinyDB simple JSON files SQLite for durable, high-perf reads with WAL SQLIte shared for high write apps DuckDB + Parquet for analytics apps Postgres + MariaDB for multi-machine client/server Great for teaching, examples, and simple deployments Amazing story of paired AI development Will completely run talkpython.fm after weeks of shared work together (in SQLite mode). Calvin #3: Jev: what to know What it is: Jev is a model from TypeSafe AI that answers with typed results (yes/no probabilities, scores, picks from your options) instead of prose. Real Python published a hands-on tutorial on 2026-09-24 and the buzz on hacker news is almost deafening. It's proprietary: Jev is a hosted, closed-weight model. There are no weights to download and no self-hosting. Everything called "open Jev" is an independent reimplementation, not TypeSafe's model. Your data leaves your machine: Every call sends your input text to a third-party API. In the tutorial that path goes through OpenRouter to TypeSafe. Think twice before sending customer messages, tickets or anything sensitive. Cost and stability are open questions: The tutorial calls Jev "cheap, but not free" and says it's fast and cheap "at the moment." It also says whether that stays true is "something to keep an eye on." Credit to Real Python: It's a good, practical intro. It shows the Noul, Score and Choice primitives, and its point that instruction wording matters more than thresholds is useful advice for any model. The tutorial itself says similar results are possible with a well-prompted LLM. Open options to look at instead: JevK5 (https://github.com/allebee/jevk5): Apache-2.0 weights and code, 4B or 9B parameters, and it accepts TypeSafe-style requests. SemIf, formerly OpenJev (https://github.com/TheoLeeCJ/openjev): MIT-licensed, small models, and it can run CPU-only. openjev-sglang (https://github.com/ekzhang/openjev-sglang): a Jev-compatible endpoint running Qwen3.6-35B-A3B, but no license is stated, so check before commercial use. The catch: These copy Jev's interface, not its model or training. Results will differ, and I haven't run any of them. Benchmarks are self-reported, and JevK5 is English-only. Michael #4: One innocent dict read makes attribute access permanently slower Timofei Ivankov benchmarks a CPython internals surprise: since 3.11, attribute access skips the instance dict entirely. A specialized opcode reads the attri.bute at a fixed byte offset in the object's inline values array. Read obj.__dict__ once, though, and the dict gets materialized, the object loses that specialized path for the rest of its life, and a million-iteration loop goes from 33 ms to 51 ms on CPython 3.14. vars() and copy.copy() trigger the same thing, so a debugging print or a shallow copy in code touching your hot objects quietly makes every later attribute access roughly 1.5x slower. The slowdown is permanent and nothing about it looks like a performance decision: ordinary code far from the hot loop can trigger it, and the function that gets slower never changes. Materializing dict produces a split table, and the LOAD_ATTR_WITH_HINT fallback declines split tables, so the object ends up with no specialization at all vars(), 'x' in o.dict, and copy.copy() all materialize it; copy.copy is the realistic trap since nobody treats a shallow copy as a performance decision slots instances read attributes at exactly the same speed and cannot fall into the trap since there is no dict to materialize On the free-threaded build both effects grow: atomic incref on reads plus an object lock on writes push the penalty from 17.6 to 25.4 ns Credit: this item was surfaced by the PyCoder's Weekly newsletter Extras Calvin: whatsnewt - a TUI text adventure through what's new in Python 3.15; playful but niche. Joke: Shipping a button in 2026…

  2. hace 6 días

    #497: Faster than light profiling

    Topics covered in this episode: Tachyon: A sampling profiler ships in Python 3.15's stdlib Python Workers are now generally available on Cloudflare Flet 1.0 - build cross-platform apps in Python marimo-book: Build static books from marimo notebooks Extras Joke Watch on YouTube Sponsored by Logfire from Pydantic: pythonbytes.fm/logfire Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedIn Calvin: Mastodon / BlueSky / X / LinkedIn Show: 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: Tachyon: A sampling profiler ships in Python 3.15's stdlib Python 3.15 adds the profiling package per PEP 799: profiling.tracing (where cProfile moved) and profiling.sampling, the new sampler called Tachyon py-spy and Austin exist but copy raw interpreter bytes with no API, so every CPython release risks breaking them; one in the stdlib is a contract to stop breaking profilers Defaults: 1 kHz, main thread, wall clock, and a -live top-like view for poking at a slow server Output is flexible: pstats, -flamegraph, -diff-flamegraph against a baseline, -heatmap on source lines, -opcodes for specialized bytecode, -gecko for Firefox Profiler with GIL and GC markers Profiling modes: wall, cpu, gil (which function is starving my other threads?), and exception, plus -async-aware to see the task graph instead of just select(), -all-threads, and -subprocesses forking a profiler per child Near-zero overhead for production; guidance is 10-30 second windows on representative load, and free-threaded builds divide the rate by thread count Attach to a running PID, same minor version only; ptrace permissions are the main friction. A 3.14 backport already exists on GitHub Caveat: it only sees Python frames, so 90% in calculate() hides NumPy underneath. For native stacks there's Cronon from HRT, 200k samples/sec over DWARF, not yet open source Calvin #2: Python Workers are now generally available on Cloudflare Python Workers are out of beta - now GA, "first-class" language on Cloudflare's Developer Platform No more manual JS interop: bindings (queues, R2, D1, Durable Objects) now work natively in Python, e.g. self.env.QUEUE.send({...}) Runs on Pyodide (WASM-compiled Python), with real TCP socket support for DB connectivity Frameworks supported: FastAPI, Django, Flask; AI libs like OpenAI SDK, LangChain, MCP Underlying platform work formalized as PEP 783 (PyEmscripten), after a year of discussion Bottom line: write real Python on Cloudflare's edge, no JS glue code required Calvin #3: Flet 1.0 - build cross-platform apps in Python Flet hits 1.0 - build Flutter-backed apps from pure Python, no frontend experience needed One codebase targets six platforms: iOS, Android, Windows, macOS, Linux, web 150+ built-in UI controls, plus support for custom controls / wrapping Flutter packages Mobile now supports real Python packages: NumPy, pandas, Pillow, cryptography Comes with pytest-based UI testing and an MCP integration for AI coding assistants Milestone lands 4+ years after its first PyPI release (Sept 2022) - signals "production ready," not experimental Michael #4: marimo-book: Build static books from marimo notebooks marimo-book is a Jupyter-Book-style static site generator built specifically for marimo .py notebooks. It ships polished multi-page sites with Material for MkDocs theming, full-text search, dark mode, and code copy, plus a content-hashed incremental build cache that drops rebuilds from 100+ seconds to roughly 3 seconds on real books. Standout extras include anywidget rendering without a kernel, static reactivity for discrete sliders via pre-rendered lookup tables, an opt-in WASM/Pyodide mode per chapter, and per-chapter launch buttons. If you've wanted to publish a marimo notebook as a real book or course site without hosting a kernel, marimo-book gives you the static, searchable, fast-loading output you'd expect from Jupyter Book. Alpha (0.1.x), but in production: pin marimo-book>=0.1.5,0.2; the book.yml schema is stable for v0.1, and dartbrains.org is a real-world user. Two-stage build by design: a marimo-aware preprocessor emits plain Markdown + inline HTML, then mkdocs (Material today, zensical tomorrow) renders it. Not a mkdocs plugin, so the shell stays swappable. Interactive widgets without a kernel: anywidget Canvas/Three.js/Plotly mounts render statically, and mo.ui.slider with explicit steps gets pre-computed as a static lookup table. WASM escape hatch per chapter: set mode: wasm and the chapter routes through marimo's MarimoIslandGenerator, shipping the marimo runtime + Pyodide bundle for full reactivity where you need it. Per-chapter launch buttons and extras: readers can jump to molab, GitHub, or a downloaded .py; optional [social], [linkcheck], and [pdf] extras cover OG cards, htmlproofer, and WeasyPrint PDF export. Sandboxed notebooks: the sandbox mode reads PEP 723 inline metadata and provisions per-notebook envs via uv for portable builds, at the cost of slower first runs. Extras Calvin: Great overview of a new feature in Python 3.15 - frozendict Michael: Microsoft Plugs Nearly 1,000 Security Holes in Windows I’ll be speaking at PyBay 2026 PyCon NL on October 15 in Utrecht Heading to Europe in October? PyCon NL is October 15th in Utrecht. One day, three tracks, about 350 people. It's an hour by train from Schiphol. There's also a session just for community organizers from groups like PyLadies, PyData, and Django. And the location fits. The Netherlands is where Python itself was born. Joke: You have homework (no really ;) ) Watch Interview with Big Data engineer in 2026 by Kai Lentit

  3. 15 sep

    #496: A lake house in Seattle

    Topics covered in this episode: Pandas Should Go Extinct Pydantic-pint puts real-world units in your Pydantic models How Libraries Run Rust Inside Python (With PyO3) AWS acquires DuckLabs Extras Joke Watch on YouTube Sponsored by Logfire from Pydantic: pythonbytes.fm/logfire Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedIn Calvin: Mastodon / BlueSky / X / LinkedIn Show: 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 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: Pandas Should Go Extinct Pandas' slowness pushes teams toward "Big Data" tools (Spark, Databricks) they don't actually need — most workloads never hit true Big Data scale Amazon Redshift telemetry: ~95% of tables are under 100GB, ~87% of queries touch 80GB or less — that's "Medium Data," not Big Data Polars and DuckDB fill that gap: single-machine, fast, no cluster required 1 Billion Row Challenge benchmark: Pandas took 4m28s vs. Polars 5.04s and DuckDB 5.19s — DuckDB also used 19x less memory On a real-world NYC taxi dataset (3GB parquet), pure DuckDB ran 2x faster than pure Pandas while using a fraction of the RAM Bonus: Apache Arrow lets you pass data between Pandas/Polars/DuckDB with zero copying, so trying them out doesn't mean a full rewrite Michael #2: Pydantic-pint puts real-world units in your Pydantic models Pydantic-pint bridges Pydantic and Pint so models can validate physical quantities like 4m or 12 meters instead of bare floats. Fields annotated with PydanticPintQuantity parse user input, convert between compatible units, and serialize quantities back out as strings. That closes a real gap for anything consuming API payloads, config files, or sensor data with measurements, letting you enforce units at the validation boundary instead of hoping every caller remembered them. via PyCoder's Weekly newsletter Unit mix-ups have literally crashed spacecraft; now your Pydantic models can refuse them at the door. Annotate a field as Annotated[Quantity, PydanticPintQuantity('km')] and inputs like 12 meters arrive auto-converted to kilometers Validation covers string, numeric, and quantity inputs, and model_dump_json serializes quantities as readable unit strings Installable from PyPI as pydantic-pint, MIT licensed, with docs at pydantic-pint.readthedocs.io Early-stage solo project at version 0.4, so API stability and maintenance are open questions worth discussing Calvin #3: How Libraries Run Rust Inside Python (With PyO3) Pydantic v2's validation core (pydantic-core) is Rust under the hood, built with PyO3 — this post shows how that bridge actually works via a small hand-built JSON parser Four steps to get Rust into Python: write a normal Rust module, annotate with PyO3 macros (#[pyfunction], #[pymodule]), compile/install with maturin, then just import it The parser builds a Rust tree first — Python never touches it until the boundary crossing Key insight: converting the Rust result into Python objects (.into_pyobject) is often the expensive part, not the parsing — 100,000 JSON values means ~100,000 Python objects built after parsing's already done Errors cross the boundary too: Rust's typed errors convert into real Python exceptions (ValueError, FileNotFoundError) via From/?, so callers get clean Python semantics Takeaway for anyone porting Rust in: if you're returning a scalar, don't sweat it; if you're returning a big structure, profile the boundary — that's the real cost, not the algorithm Michael #4: AWS acquires DuckLabs Thank you Dylan McConnell. What does this mean for the DuckDB ecosystem? DuckDB is the open-source in-process analytical SQL engine. MIT licensed. The IP is not owned by any company - it's held by the nonprofit DuckDB Foundation, which was created when the team spun out of CWI Amsterdam. Peter Boncz, the CWI representative on the Foundation board, describes it as the entity that holds all IP of open-source DuckDB. DuckLabs (ducklabs.com) is the company, formerly branded DuckDB Labs. Founded a little over five years ago by Hannes Mühleisen and Mark Raasveldt to give the DuckDB team a stable long-term home, bootstrapped deliberately instead of taking VC, grown to 30+ people in Amsterdam, funded by support and feature-prioritization contracts. It employs the core devs. It does not own DuckDB. DuckLake is one of three projects DuckLabs builds, what they call the Duck Stack: DuckDB, DuckLake, and Quack. DuckLake is the lakehouse format that puts catalog metadata in a SQL database instead of in files on object storage. Quack is newer - an RPC-style protocol that turns DuckDB into a client-server system where both ends are DuckDB instances, slated to stabilize in DuckDB v2.0 in September 2026. MotherDuck is a separate Seattle company, Jordan Tigani's, selling serverless hosted DuckDB. It was started in partnership with DuckDB Labs and has worked closely with Hannes and Mark for four years. It contracted DuckLabs for engineering work and contributes heavily upstream - three of its engineers are among the top 10 outside contributors to DuckDB. It also sells its own DuckLake offering. Customer and collaborator, never owner. What the AWS post changes. Amazon bought the company, not the project. DuckLabs joined AWS effective September 1, with the process concluding August 31, 2026. Hannes and Mark keep leading the team and the project's technical direction, the team stays in Amsterdam, and DuckDB stays MIT under the Foundation. AWS gets the people and a direct line to the roadmap. The license protects your code, not your priorities. Three second-order effects worth tracking: The Foundation board is the real question. It has three directors: Mühleisen, Raasveldt, and Boncz. Two now work for AWS. Commentary on the deal has focused on exactly this - the license protects the code, not the roadmap. The announced counterweight is governance: a technical advisory board on the Foundation, and opening the extension stack so extensions signed by other developers can run in DuckDB. MotherDuck immediately moved into the business DuckLabs vacated. It now sells DuckDB enterprise support, which it had avoided because it didn't want to compete with DuckLabs' business model, and says it has explicit blessing from Hannes and Mark now that they're joining Amazon. It also bought Tower.dev the day before the AWS announcement. Everyone expects an AWS DuckDB service. Tigani says Amazon will likely release one eventually, and welcomes the competition, citing Redshift's failure to slow Snowflake on AWS. The groundwork is already visible: Amazon Quick uses DuckDB to query S3 Tables and has processed over 2.5B queries with it since launching in October 2025. The DuckLake angle is the one to watch. AWS is heavily committed to Iceberg through S3 Tables, and it just acquired the team behind a competing lakehouse format. The stated plan is to use DuckDB, DuckLake, and Quack together to power a new generation of data services, but which format wins internal priority is unannounced. Extras Calvin: astral-sh/uv 0.12.12: code-signed release binaries 🥳 Michael: My MacBook power supply rebooted to install updates (?!?) The Story of VS Code | Official Documentary Amazon/AWS acquires DuckLabs (see recent episode on DuckLake) Joke: We’re agentic now

  4. 8 sep

    #495: Banned

    Topics covered in this episode: EuroPython 2026 videos are online The State of Django 2026: Boring is so back htmx 4.0.0 has been released 🐍 Functionally Zen Extras Joke Watch on YouTube Sponsored by us! Support our work through: Our courses at Talk Python Consulting from Six Feet Up Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedIn Calvin: Mastodon / BlueSky / X / LinkedIn Show: 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 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: EuroPython 2026 videos are online The EuroPython Society has published all 117 recordings from EuroPython 2026 on the official EuroPython Conference YouTube channel. The conference ran July 13-19 in Krakow, Poland and celebrated the conference series' 25th anniversary. The playlist covers keynotes, panels, lightning talks, and full talk recordings across Python core, web, DevOps, data/ML, embedded, and other tracks. If you missed EuroPython 2026 in Krakow, this is the complete free on-demand archive of one of the year's biggest European Python events. 117 videos now live on the EuroPython Conference YouTube channel, last updated Aug 17, 2026. Michael’s personal watch list. Calvin #2: The State of Django 2026: Boring is so back State of Django 2026 (JetBrains/DSF survey, ~3,500 devs, 40+ countries) - "boring is so back": Django's core stays reliable while everything around it moves fast Core is stable: Postgres 76–79% for 5 years running, templates ~80%, 43% already on Django 6.0 AI is routine now (only 10% use none) but workflow's unsettled - Claude Code leads at 35%, and 56% still just use it for chat, not autonomous edits Tooling is consolidating: uv and Ruff both at 43% adoption, each replacing several older single-purpose tools Type hints are winning (57% use them) but the checker is up for grabs - IDE-built-in leads at 40%, Mypy 32%, with ty/Pyrefly emerging Two Django communities coexist happily: 72% server-rendered templates vs. 53% API-only - and htmx adoption jumped from 5% to 34% in five years Michael #3: htmx 4.0.0 has been released After 8 months of work, the htmx team shipped 4.0.0, a rewrite that moves internals from XMLHttpRequest to fetch() while keeping the API almost identical to htmx 2. Three changes may need action: attribute inheritance is now explicit via an :inherited suffix, event names follow a htmx:phase:action pattern, and history no longer caches pages in localStorage. Additions include built-in morph swaps, the new hx-partial tag, and many core extensions. htmx 2 stays supported and remains latest on npm until early 2027. htmx is the go-to frontend layer for Python server-rendered apps (Flask, Django, FastAPI), and 4.0 is deliberately low-drama: nearly behavior-compatible, so teams can upgrade on their own schedule and pick up morph swaps and streaming extensions. Explicit inheritance is the biggest migration item: hx-confirm, hx-headers, hx-target and friends no longer cascade to children unless you append :inherited; hx-disinherit and hx-inherit are gone A CLI upgrade checker (npx htmx.org@4.0.0 upgrade-check) flags spots needing :inherited, renames like hx-disable to hx-ignore, removed attrs like hx-vars, and old event names in templates and JS Events follow htmx:phase:action (htmx:beforeRequest becomes htmx:before:request); most error events collapse into htmx:error and htmx:xhr:* events are removed with XMLHttpRequest History no longer snapshots pages in localStorage; back navigation re-fetches and swaps into the body, fixing a chronic support headache, with a new hx-history-cache extension for sessionStorage caching New features: out-of-the-box morphing swaps, the [HTML_REMOVED] tag for multi-element updates, streaming over SSE/WebSockets/multipart, and hx-live, their Alpine-inspired DOM scripting solution No forced upgrade: 2.x stays latest on npm until early 2027 (4.0 remains next) and is supported indefinitely; the team even ships official LLM skill files for guidance and upgrading Calvin #4: 🐍 Functionally Zen Functionally Zen (Kyle Adams, Test Double) - riffs on "simple is better than complex" with 7 extra tenets for Python simplicity Core claims: idiomatic > non-idiomatic, data > functions, pure functions > impure functions > classes Favorite example: a medical-dosage calculator replaced with a plain lookup dict - no logic, no tests needed Big idea: keep a thin "impure shell" around a "pure core" (Gary Bernhardt's functional core / imperative shell) - push side effects (API calls, DB, files) to the edges Side note: constructors that do I/O are "poison pills" - the side effect infects every class that depends on them Payoff: pure functions and no-mock tests are just easier to read and reason about than the alternative Extras Calvin: uv ships trusted-publisher token revocation and Python 3.15 support Making a Python interpreter in 1024 bytes Michael: Steering council voting is now open Joke: Makes you look like this?

  5. 1 sep

    #494: Python Wrapture

    Topics covered in this episode: OpenAI's Python SDK has migrated to HTTPX2 TMOG - Native Task Manager for macOS, Windows, and Linux wrapture - one wrapper for mocking, tracing, and observability linkedin2md: turn your LinkedIn export into 40+ Markdown files Extras Joke Watch on YouTube About the show Sponsored by us! Support our work through: Our courses at Talk Python Consulting from Six Feet Up Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedIn Calvin: Mastodon / BlueSky / X / LinkedIn Show: 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 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: OpenAI's Python SDK has migrated to HTTPX2 The OpenAI Python SDK has migrated to HTTPX2, the Pydantic-stewarded fork of httpx. Pydantic picked it up citing "limited activity recently" in the original project, promising "a reliably maintained path forward." If you just use the default client, nothing to do. No code changes. The catch is TLS. Quoting the guide: HTTPX "previously verified certificates against the CA bundle provided by certifi. HTTPX2 instead uses the operating-system trust store, and the SDK no longer installs certifi." That "can break certificate verification in minimal container images without system CA certificates, environments using corporate TLS-inspecting proxies, and deployments that relied on a custom or modified certifi bundle." The fix is SSL_CERT_FILE or SSL_CERT_DIR, or pass your own ssl.SSLContext via verify. Deeper integrations need real edits: custom clients, auth handlers, hooks, and request mocking all take HTTPX2 objects now, and plain httpx is no longer pulled in transitively. So import httpx in your own code means declaring it yourself or moving over. Temporary escape hatch: a legacy HTTPX client Michael #2: TMOG - Native Task Manager for macOS, Windows, and Linux A native, deeply instrumented system monitor for macOS, Windows, and Linux, now in public beta - from Plummers' Software, i.e. Dave Plummer, who wrote the original Windows Task Manager and donated it to Microsoft in 1995. Wikipedia Three real native apps: Swift/AppKit on macOS, Win32 on Windows, C++/Qt 6 on Linux, with a shared C++ core keeping metric semantics aligned - no browser shell anywhere. One dense summary: CPU, clocks, thermals, GPU, memory, storage, network, energy, and the processes responsible for the load, all click-through. Per-core honesty: logical processor and NUMA views, P and E cores color-coded, optional kernel time, 60 FPS live meters. Memory with context: pressure, wired, compressed, cached, committed, available, and swap, plus configurable scrolling history. Processes that act like processes: tree view, filtering, sorting, follow mode, and native verbs including service and launchd control. Phosphor themes: light, dark, green, amber, blue, or mono, with color and saturation you tune yourself. Calvin #3: wrapture - one wrapper for mocking, tracing, and observability Graham Dumpleton, author of wrapt and the original New Relic Python agent, has released wrapture. The name is wrapt plus capture. The core idea: wrap real code instead of replacing it, so the real code still runs while you watch every call. Name a method with wrapture.binding(Class, "method"), open a timeline(), and you get a tape of what actually happened. Real return values, real nesting, arguments normalised against real signatures. tape.tree() prints the call graph as it ran. One mechanism, three jobs: monkey patching with a real lifecycle (apply, remove, suspend, plus returns, raises, transforms_args), unit testing that asserts on real call flow instead of a flat MagicMock call list, and ad-hoc tracing of a running app. The testing pitch is error paths. Inject TimeoutError at the payment gateway, then assert the ledger was never written. Stubs and mocks are strict and spec-required, and there is deliberately no bare Mock(). Tracing needs no code at all. A wrapture.toml naming targets and a sink, run with python -m wrapture main.py, and you get a live call tree with timings. It captures ordinary logging calls as nested events, and with the otel extra it exports spans, metrics and correlated logs with W3C trace ids that join across services. Every line of code and docs was AI-written under their direction, and they say so up front. Two weeks from first commit, eleventh alpha, over 1000 tests, 150+ pages of docs. Alpha on PyPI, needs Python 3.12+ and wrapt 2.4.0+. Michael #4: linkedin2md: turn your LinkedIn export into 40+ Markdown files Via Juan Manuel Daza - a Python CLI that unpacks LinkedIn's data-export ZIP into clean, per-category Markdown you can drop straight into an LLM. One command: linkedin2md Complete_LinkedInDataExport.zip, plus o for output dir, -lang en|es, and -pdf. 40+ output files: profile, experience, education, skills, connections, posts, comments, reactions, recommendations, endorsements, job applications, even ad targeting and LinkedIn's inferences about you. Built for LLM analysis: the README pitches NotebookLM, Claude Projects, Obsidian, and Ollama, with example prompts like "what patterns do you see in my career transitions?" PDF resume mode: -pdf renders an A4 CV via weasyprint, and degrades gracefully to Markdown-only if it isn't installed. Dependency note: "pure Python / zero-dep" holds for the Markdown path only - the PDF path needs weasyprint and markdown installed. Install: pipx install linkedin2md recommended, pip in a venv otherwise - 86% Python, 10 releases, v0.3.1 in May. Agentic dev angle: repo ships opencode config and an N3RV subagent pipeline, including a "judgment day" dual-model adversarial PR review. Extras Calvin: EVE Online Migrates to Python 3 Michael: Dinkus by Will McGugan Joke: Tao of Programming: Book 5 Maintenance

  6. 26 ago

    #493: CalVer and LTS

    Topics covered in this episode: Web UIs for your reverse proxy Wagtail 8.0 is hot off the presses RISC-V is now officially supported by CPython Django’s annual releases make every version an LTS Extras Joke Watch on YouTube About the show Sponsored by Logfire from Pydantic: pythonbytes.fm/logfire Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedIn Calvin: Mastodon / BlueSky / X / LinkedIn Show: 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: Web UIs for your reverse proxy Traefik, nginx, and Caddy all sit in front of a lot of self-hosted infrastructure, and all three are configured by hand-editing files. Three active projects put a control plane on top: Traefik Manager (Python + Flask), Nginx UI (Go + Vue), and caddy/ui (React + Node). All three are additive rather than replacements - none of them take ownership of your config away from you - which is the part that matters when the thing has write access to production routing. Traefik Manager is the Python one: Flask 3.1 and Gunicorn for the control plane, a lightweight Go agent for remote instances, currently v1.10.0 with an Android companion app. Nginx UI is a single Go binary at 11.3k stars, with a block-style config editor, an Ace editor doing LLM completion on nginx syntax, and an MCP server so agents can drive it. caddy/ui runs as two containers next to your existing Caddy, reads and writes your Caddyfile directly, and uses Caddy's /adapt API to validate before reload - no Docker socket required. Each one edits the config the underlying server already reads, so your files stay the source of truth and you can drop the UI without unwinding anything. Undo is a first-class feature across all three - timestamped backups with optional Git history, config version compare and restore, Caddyfile snapshots with one-click rollback. Observability is where they diverge: Traefik Manager does CrowdSec and a visual route map, Nginx UI does server metrics, caddy/ui streams access logs over SSE and pulls p50/p95/p99 off Caddy's Prometheus endpoint. Maturity spread is wide - Nginx UI has 11.3k stars, caddy/ui has 4 and was built in a single Claude session - and caddy/ui ships with auth off by default, so set CADDY_UI_USER and JWT_SECRET before it goes anywhere near a public interface. Calvin #2: Wagtail 8.0 is hot off the presses Link: https://github.com/wagtail/wagtail/releases/tag/v8.0 Custom base page models are now supported, so projects aren't locked into subclassing Wagtail's Page as shipped (Matt Westcott). New v3 REST API handles both read and write CMS operations, a first for Wagtail's API. A global registry for permission policies, plus full customizability for the remaining page views via PageViewSet. AVIF and WebP images are no longer auto-converted to PNG by default, a real behavior change to watch on upgrade. Five security fixes: page admin API restrictions, document identification by SHA1 hash, descendant collections in the Documents/Images API, snippet copy permissions, and the page translation endpoint. Formalized Django 6.1 support, and CI now runs on uv with a lockfile. Sponsor: Logfire from Pydantic Your AI agent failed at 2am. Was it the model? A tool call? The database? Most observability tools can't tell you, because they only see part of your stack. Pydantic Logfire sees all of it. One trace across your agents, LLMs, APIs, and database. Down to the infrastructure: services, Kubernetes, and hosts. It's built on OpenTelemetry, with SDKs for Python, TypeScript, and Rust, and it works with any OTel-compatible language. Every prompt, token count, and cost, right next to your vector searches and API calls. You query everything with Postgres-compatible SQL. And so can your coding agent, through the Logfire MCP server. Stop guessing. Read the trace. Pydantic Logfire. AI, it's still just engineering. Visit pythonbytes.fm/logfire today and sign up today. Get 10M records free every month, no card required. You can even click “Onboard with your coding agent” to copy a prompt to have claude or codex integrate Logfire into your app. Thanks to Pydantic for supporting the show. Calvin #3: RISC-V is now officially supported by CPython Link: https://blog.python.org/2026/08/riscv-now-officially-supported/ CPython added RISC-V as a tier 3 platform under PEP 11, specifically the 64-bit Linux target riscv64-unknown-linux-gnu. RISC-V is an open ISA anyone can implement, unlike x86 and ARM, and its market is projected to quadruple by 2032. The RISE Project donated real RISC-V machines for buildbots; the author's work was funded by a Sovereign Tech Agency fellowship. What changes: the port is now a maintained compatibility target, so CPython changes are less likely to quietly break it. What doesn't: no python.org installers, no binary wheel parity for native extensions. Next up: RISC-V runners in CPython CI for pre-merge feedback, then a push toward tier 2, plus architecture-specific optimizations. The ask is testing. If you have RISC-V hardware, build CPython, run your test suite, file what breaks. Tier 3 is the weakest support tier. PEP 11 tier 3 requires a core developer contact and a buildbot, but failures on tier 3 platforms explicitly do not block a release. Saying "ongoing CI/testing expectations" oversells it. The honest bit is "someone is now on the hook for it, and breakage gets noticed," not "it's guaranteed working." Worth the caveat that this is Linux SBCs, not microcontrollers. A VisionFive 2 counts, an ESP32-C6 or Pico 2 does not. Those are 32-bit non-Linux parts where MicroPython is still the answer. Michael #4: Django’s annual releases make every version an LTS Starting with Django 2028, Django will move to one January feature release per year, adopt calendar-based version numbers, and support every release for three years. The old distinction between standard and LTS releases disappears, giving teams a predictable annual upgrade path that aligns more closely with Python’s own release and support cadence. Every Django release becomes the safe, long-supported choice, so teams no longer need to wait for a specially designated LTS version or absorb two years of changes at once. Each release gets one year of mainstream bug fixes followed by two years of security and data-loss fixes. New releases support the three latest Python versions and add the next Python release during their first year. Calendar versioning begins with Django 2028, followed by Django 2029 and so on. Three Django versions will be supported at any time, giving third-party packages a clearer rolling target. Nothing changes before 2028, and existing commitments for Django 5.2 LTS and 6.2 LTS remain in place. Extras Calvin: The Python docs now document the time complexity of built-in types https://docs.python.org/3.16/library/time-complexity.html Thinking in Python - Bruce Eckel's free book https://thinkinginpython.com/ Michael: prune_uv_pythons.py - Prune uv-managed Python installs, keeping only the newest patch per minor version Runs automatically in my system “upgrade” script: upgrade-output-2026.png Started using Ollama cloud models for my Hermes assistant. Thanks to Jeff Triplett I learned they are not just local models. Joke: The Tao of Programming - Book Seven: Corporate Wisdom

  7. 18 ago

    #492: Codeberg Puts Head in Sand

    Topics covered in this episode: Python 3.12.14, 3.11.16, 3.10.21 - security releases Codeberg’s AI-code ban tests its role as a GitHub alternative Brett Cannon: what's missing for reproducible builds on PyPI nothing records the source code a distribution came from. direct_url.json captures it when you install from a repo or archive, so the fix is putting the same info in sdist/wheel metadata. recording the build tools. Wheels can already do this via PEP 770 SBOMs in .dist-info/sboms/ - sdists can't, since they're a tarball plus a precalculated PKG-INFO with nowhere to hang extra metadata. Either "don't use sdists" or an sdist v2. Extra extra extra, hear all about it Extras Joke Watch on YouTube Sponsored by Logfire from Pydantic pythonbytes.fm/logfire This episode is brought to you by Pydantic Logfire. It's observability for AI apps from the team behind Pydantic - agents, LLMs, APIs, database, and infrastructure in a single trace, queried with Postgres-compatible SQL. Your coding agent can query it too, through their MCP server. I'll tell you more later. Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedIn Calvin: Mastodon / BlueSky / X / LinkedIn Show: 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: Python 3.12.14, 3.11.16, 3.10.21 - security releases https://blog.python.org/2026/08/python-31214-31116-31021/ Source-only security releases for the three branches now in security-fix-only mode; release team blamed the European solar eclipse for the timing. tarfile hardening. Multiple path-traversal bypasses of the data filter closed, including a symlink escape that bypassed the CVE-2025-4330 fix; extract() now applies the filter to link targets too. Four fresh CVEs: CVE-2026-2297 (SourcelessFileLoader not using io.open_code() for .pyc), CVE-2026-4224 (expat crash on deeply nested content models), CVE-2026-3644 (control chars in http.cookies.Morsel), plus the completed CVE-2021-4189 fix in ftplib.ftpcp. Quadratic-complexity DoS cleanup across the stdlib: HTMLParser, configparser regexes, unicodedata.normalize(), csv.Sniffer.sniff(), and ElementTree XPath index predicates. Header/injection fixes: CR/LF rejected in HTTPConnection.set_tunnel(), control chars blocked in wsgiref.handlers status, and webbrowser now rejects leading dashes (plus a %action prefix bypass). http.client now caps chunked trailer lines and 1xx interim responses at 100 each - a hostile server could previously hang the client forever despite a socket timeout. Memory-safety odds and ends: stale pointers in lzma/bz2/zlib decompressors after MemoryError, a bz2 stack overflow on reuse-after-error, and bundled libexpat bumped to 2.8.3. If you're still on 3.10, 3.11, or 3.12 - and you extract tarballs from anywhere you don't fully control - this one's not optional. Michael #2: Codeberg’s AI-code ban tests its role as a GitHub alternative Armin’s article “Codeberg Divides” Armin Ronacher argues that Codeberg’s new terms, which prohibit projects mostly written with generative AI, create a vague and difficult-to-enforce boundary. His larger concern is that a democratically governed host can still be unpredictable or ideologically narrow, weakening Codeberg’s potential as a broad European alternative to GitHub. The strongest question for Python developers is whether repository hosting should judge legal open source by how code was produced, or focus on behavior and resource abuse. “Mostly generated” is hard to measure in modern codebases where developers mix handwritten code, completions, agents, and generated refactors. Ronacher suggests clearer alternatives: ban all LLM involvement, or target autonomous repository spam, abusive resource use, and low-quality generated contributions directly. Codeberg is free to choose a values-driven community, but that may conflict with being predictable, neutral infrastructure and a serious GitHub competitor. Worth discussing: can open-source communities set meaningful AI boundaries without driving maintainers and projects into opposing camps? Very first search for these terms lands on this page. Codeberg looked like a viable alternative. … Unfortunately, the latest update to its terms of service seems to mark a first step in changing one part I moved there for, namely the “freedom” part. Sponsor: Logfire from Pydantic Your AI agent failed at 2am. Was it the model? A tool call? The database? Most observability tools can't tell you, because they only see part of your stack. Pydantic Logfire sees all of it. One trace across your agents, LLMs, APIs, and database. Down to the infrastructure: services, Kubernetes, and hosts. It's built on OpenTelemetry, with SDKs for Python, TypeScript, and Rust, and it works with any OTel-compatible language. Every prompt, token count, and cost, right next to your vector searches and API calls. You query everything with Postgres-compatible SQL. And so can your coding agent, through the Logfire MCP server. Stop guessing. Read the trace. Pydantic Logfire. AI, it's still just engineering. Visit pythonbytes.fm/logfire today and sign up today. Get 10M records free every month, no card required. You can even click “Onboard with your coding agent” to copy a prompt to have claude or codex integrate Logfire into your app. Thanks to Pydantic for supporting the show. Calvin #3: Brett Cannon: what's missing for reproducible builds on PyPI Framing came out of his 2026 Python Packaging Council nomination - the secure-supply-chain gap he found is that Python has no defined way to do reproducible builds at all. Design goal is zero friction: producers uploading to PyPI shouldn't have to do anything. The work lands on build backends and installers. Gap #1: nothing records the source code a distribution came from. direct_url.json captures it when you install from a repo or archive, so the fix is putting the same info in sdist/wheel metadata. Gap #2: recording the build tools. Wheels can already do this via PEP 770 SBOMs in .dist-info/sboms/ - sdists can't, since they're a tarball plus a precalculated PKG-INFO with nowhere to hang extra metadata. Either "don't use sdists" or an sdist v2. The replay mechanism already exists: [build-system] in pyproject.toml is a defined entry point, so if backends recorded their own environment, you could reinstall and re-run the build. Payoff idea: trusted third parties report successful reproductions back to PyPI, which displays "independently reproduced by X" - surfaced in the index API so installers could prefer reproduced files. Explicitly framed as a perk, not a requirement - roughly SLSA build level 1, no shaming projects that don't opt in. Verbal kicker option: "And don't think pure-Python wheels are off the hook. Something built that wheel, and if that something was compromised, so is your wheel. SolarWinds was a build-process attack." Michael #4: Extra extra extra, hear all about it Python 3.14.7 Upgraded the MCP servers to 2026-07-28 v2 protocols (talk python, python bytes) Got agentsview running synced via postgres Talk Python courses, teams trial offering Talk Python courses, government procurement offering Lean TDD audio book is out Extras Calvin: uv now prefers post-quantum key exchange - https://github.com/astral-sh/uv/releases/tag/0.12.4 Joke: Beware of dog

  8. 12 ago

    #491: Feeling Judged

    Topics covered in this episode: Claude Code /insights Post-quantum crypto lands in Python MCP goes stateless — and FastMCP gets renamed inshellisense - IDE style command line auto complete Extras Joke Watch on YouTube About the show Sponsored by Xweather Xweather combines enterprise-grade weather intelligence with agent-ready APIs, natural language capabilities, and an MCP server so your agents can adapt workflows, automate responses, and make better decisions based on real-world conditions. Michael will tell you more about them later in the show. Get started for free at pythonbytes.fm/xweather Connect with the hosts Michael: Mastodon / BlueSky / X / LinkedIn Calvin: Mastodon / BlueSky / X / LinkedIn Show: 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 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: Claude Code /insights Michael’s Insights: michael-kennedy-claude-code-insights-2026-08-09.html Be careful sharing these outputs, they include details references to your projects, errors, security findings, etc. ;) /insights reads your last 30 days of local session transcripts and hands back an interactive HTML report on how you actually work. One command, zero setup: type /insights in a session, or run claude -p "/insights" from the shell for a non-interactive version that just prints the path Reads what's already on disk: pulls session logs from ~/.claude/projects/, skipping agent sub-sessions and anything under 2 messages or 1 minute Project areas: clusters your sessions into themes like "CLI Tooling" or "Documentation" with session counts Friction analysis: categorizes where things went wrong by root cause - and quotes your own prompts back at you Interaction style: tells you whether you're a delegator or a micromanager, plus which workflows are worth doubling down on Actually actionable: suggests concrete CLAUDE.md additions and Claude Code features you're not using The catch: Haiku does the per-session classification, so the first run takes several minutes; results cache to ~/.claude/usage-data/facets/ and the report lands at ~/.claude/usage-data/report.html Calvin #2: Post-quantum crypto lands in Python pyca/cryptography 48 ships ML-KEM (key establishment) and ML-DSA (signatures) — NIST's post-quantum standards, now one pip install away. Big deal because it's the 11th most-downloaded package on PyPI (~1.2B downloads/month) and sits under Ansible, Certbot, Airflow, and paramiko. No PQ there, no PQ anywhere in Python. Trail of Bits did the work (Rust bindings, cross-backend API, tests, AWS-LC backend support), funded by the Sovereign Tech Agency. Timing tracks a June 22 White House order setting federal deadlines: PQ key establishment by end of 2030, PQ signatures by end of 2031. Not a drop-in swap — the wire sizes explode. ML-DSA-65 signatures are 3,309 bytes vs Ed25519's 64; ML-KEM-768 public keys are 1,184 bytes vs X25519's 32. Hardcoded field sizes and length prefixes will bite. API looks like the existing asymmetric primitives, except ML-KEM is encapsulate/decapsulate rather than a Diffie-Hellman exchange. SLH-DSA (the hash-based conservative backstop) is still in progress. The primitives are here, but protocols haven't caught up — so you won't be running post-quantum Certbot this week. Sponsor: Xweather You're using agents that can write code, summarize documents, and automate workflows. But they're missing one thing: awareness of the world around them. This is where today's sponsor, Xweather comes in. Xweather combines enterprise-grade weather intelligence with agent-ready APIs, natural language capabilities, and an MCP server built for tools like Claude, Codex, Copilot, and modern IDEs – so your agents can adapt workflows, automate responses, and make better decisions based on real-world conditions. Backed by Vaisala, whose instruments fly on NASA missions to Mars, Xweather delivers trusted data and unique insights that go beyond conditions to actual impact – from real-time lightning strikes to road surface forecasts. Start with 15,000 free API calls each month and pay only for what you use as you grow. Xweather is your full weather stack, for developers by developers. Start building for free today at pythonbytes.fm/xweather. The link is in your podcast player's show notes and on the episode page. Thanks so much to Xweather for supporting Python Bytes. Calvin #3: MCP goes stateless — and FastMCP gets renamed From Philipp Acsany over at Real Python The 2026-07-28 spec landed July 28 and the Python SDK shipped 2.0.0 the same day. Biggest rewrite since MCP launched, and it's breaking on purpose. Context for scale: the Tier 1 SDKs are pulling close to half a billion downloads a month, with TypeScript and Python each past a billion total. The headline is the stateless core. The initialize/initialized handshake and the Mcp-Session-Id header are both retired — protocol version, client identity, and capabilities now ride in _meta on every request, with an optional server/discover RPC if a client wants capabilities up front. Any request can land on any instance behind plain round-robin, no shared storage. Server-initiated calls are the hard part of the migration. Sampling, elicitation, and roots/list no longer call back to the client; instead the server returns resultType: "input_required" and the client retries with inputResponses attached. Multi Round-Trip Requests, MRTR. Also: Mcp-Method and Mcp-Name are now required headers so gateways route on headers instead of cracking JSON bodies, and missing-resource errors move to standard 32602. Deprecation sweep with an actual policy behind it — Roots, Sampling, Logging, and the legacy HTTP+SSE transport all deprecated with a twelve-month minimum offramp. Tasks graduated out of the experimental core into a real extension, which is what the formalized extensions framework was for. MCP Apps is now an official extension too, so a tool call can return sandboxed interactive HTML. Auth picked up RFC 9207 issuer validation, issuer-bound credentials, and a shift from DCR toward CIMD. Python SDK 2.0 is where it gets personal: FastMCP is now MCPServer, no alias, no shim. McpError → MCPError. Wire types went snake_case (is_error, input_schema) and moved to a standalone mcp_types package, with mcp.types kept as a permanent alias. One Client object replaces the old transport + ClientSession + initialize() stack. httpx became httpx2. Sync handlers run on worker threads now, so asyncio.get_running_loop() raises inside them. The good news: one MCPServer serves both protocol eras, so 2025-era clients keep working with nothing to configure, and a Resolve(fn) parameter lets one tool body cover MRTR and the old path. 1.x is maintenance-and-security-fixes only — pin mcp>=1.28,2 if this week is already full. The Tasks extension isn't in 2.0.0 yet, so Tasks has left the core spec but hasn't landed in the SDK. If you only call MCP servers, you mostly just get the benefits for free. If you ship one, you already know what your week looks like. And if you use the standalone fastmcp package instead of the official SDK — different project, 3.x line, none of this touches you. The rename is partly to stop the two from being confused. Michael #4: inshellisense - IDE style command line auto complete via Doug Nichols inshellisense provides IDE style autocomplete for shells. It's a terminal native runtime for autocomplete which has support for 600+ command line tools. inshellisense supports Windows, Linux, & macOS. If you are using a NerdFont patched font, you can enable the NerdFonts support in your config file Extras Calvin: Django 6.1 Released — https://www.djangoproject.com/weblog/2026/aug/05/django-61-released/ DjangoCon US is quickly arriving, grab your tickets now! — https://2026.djangocon.us/ Michael: AI integration: Python Bytes for AI Up and Running with Rust Course is out! Joke: But they already know

Ver todo (498)

Anfitriones e invitados

Acerca de

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.

Información

  • Creador
    Michael Kennedy and Calvin Hendryx-Parker
  • Años de actividad
    2016 - 2026
  • Episodios
    498
  • Clasificación
    Apto
  • Copyright
    © Copyright 2016-2026
  • Mostrar sitio web
    Python Bytes

También te podría interesar