Blogwalk — eleven years of kubernetes.io/blog as a walkable archive

767 blog posts, March 2015 to August 2026, censused one by one. Some of them are still exercises you can run. Most of them broke, and the ones that broke are the valuable ones.

This is supplementary. It gates no phase; P0–P12 are complete without it. It has its own reading order and links into phases/ nowhere.

Where to start

Chronologically: 2015 first, then each year in turn, exercise 01 upward inside it. The twelve years below is the index. That order is not a preference; it is the order the tree is already in. Exercise numbering is publication order, checked by the gate, and it runs the same direction as the archive's references to itself. From 2018 on, the commonest reason a post lost its walk was that an earlier year's exercise already carries its diff — four of 2018's largest announcements, all four of 2023's Gateway API posts, twenty of 2025's 78 rows. The dependencies point backward in time because backward in time is where the diffs were first walked.

Exercises stand alone and run alone, as in labs/, so the order is recommended and not enforced. An exercise may point at any exercise earlier in this spine — an earlier year, or a lower NN in the same year — and never at a later one; and such a pointer is always a pointer, never a prerequisite. Nothing here waits on an exercise you have not done, so a reader arriving in the middle can always follow a link back and will never find it unwritten.

By topic instead. The topic column of the twelve censuses is the index, and the whole mechanism is a grep:

grep '`walk` | `api`' blogwalk/*/README.md

Nine of the twelve tags carry an exercise at all — api 28, nodes 23, net and storage 17, security 15, tooling 13, sched 11, obs 10, and etcd exactly 1 — while ecosystem, history and meta carry none, the first by the vendor rule and the other two by their nature. So this is a route over nine tags of very uneven size, one of which is a single file, and it crosses the grain of the backward references above rather than running with it. The rule in the paragraph above is what makes it safe to take anyway.

Or start at 2021, if ten years of archaeology before your first exercise that runs as written is too long a wait. 2021 is where the blog reorganises around the release train — the reject rate halves to 21%, and 62% of the year's posts land in the three release months — and the 76 walks from there on are mostly release-feature exercises, where the diff is a feature gate's graduation rather than a broken command. The door is real and this file will not pretend otherwise. But taking it costs you the argument the whole tree is built on: a post that still works teaches you 2015; a post that broke teaches you what changed and why. The ones that broke are the valuable ones, and at that entrance nearly all of them are behind you.

The pin

The corpus is kubernetes/website at 7c76070faf9b19e6a417c446043dbafd10a7aa1d (2026-08-26), frozen for the whole sweep. The newest Kubernetes at that pin is v1.37 "Garhwal", released on the pin date itself. "767 posts" is a true number only against a commit, and a census gate that reads a moving target is red every Monday for reasons that have nothing to do with the census.

Re-create it:

git init website && cd website
git remote add origin https://github.com/kubernetes/website.git
git sparse-checkout init --cone
git sparse-checkout set content/en/blog/_posts content/en/docs content/en/examples
git fetch --depth 1 origin 7c76070faf9b19e6a417c446043dbafd10a7aa1d
git checkout FETCH_HEAD

The census itself reads only _posts. The other two paths are in the checkout because the exercises measure the pinned documentation and example trees against the posts, and an exercise that cites a file:line is citing that commit.

Then, from the root of this repo:

python3 blogwalk/build-manifest.py <path-to-website>/content/en/blog/_posts

That regenerates manifest.tsv — one row per post, the worklist every census row is checked back to — and asserts all twelve per-year counts on the way. If the assertion fires, the pin moved; the manifest did not drift quietly.

A post is any file under _posts whose frontmatter carries a title. Not any file ending .md: two published posts have no extension at all and take their slug from their title, one opens with a UTF-8 BOM, two open with a blank line, one opens with six dashes. 130 posts are page bundles at <slug>/index.md, over half of 2024–2025, so a glob of <year>/*.md silently misses them. path is the key, not slug — 19 rows collide on 12 slugs, six of them on weekly-kubernetes-community-hangout inside 2015 alone.

Of the 767, 756 are published; 11 are draft: true, the v1.37 feature-blog queue, still unpublished at the pin because v1.37 shipped that day. They carry a draft column rather than a deletion.

A post's URL is resolved, not derived, which is why the manifest carries a url column and why no census may build a link by hand. Only 552 posts take the site-wide permalink from hugo.toml, /:section/:year/:month/:day/:slug/. 205 carry a per-post url: override in the month-only form — all of 2015, 2016 and 2017, twelve of 2018, six of 2020 — and eleven of those are hand-written literal paths whose capitalisation is not the slug's, so weekly-kubernetes-community-hangout is served at /blog/2015/03/Weekly-Kubernetes-Community-Hangout/. Ten of the eleven drafts are undated, so nothing resolves for them and their url is empty; the eleventh is dated, so a permalink does resolve and the manifest records one the site answers with 404. 757 URLs, no collisions. Hugo's precedence — per-post override first, site template second, literal used as written — is implemented once in build-manifest.py and nowhere else.

2026 is partial as of the pin (70 posts, to 2026-08). Refreshing to a later pin is a separate, cheap job once the census exists: it is a diff against a file listing.

Then → Now: what an exercise here actually is

Nothing from 2015 runs as written. v1beta3, ReplicationControllers, no RBAC, createExternalLoadBalancer, Kubernetes 0.15. So an exercise does the post's intent on a modern cluster, and the payload is the diff and the argument behind it. strong-simple-ssl-for-kubernetes becomes cert rotation. introducing-kubernetes-v1beta3 becomes why the API is versioned at all. RC → Deployment is a live argument about who owns a rollout.

This inverts the problem: a post that still works teaches you its own year; a post that broke teaches you what changed and why. It is the move strands/source-archaeology.md makes for file paths — a path move is usually a design decision leaving a trace — applied to prose instead of paths.

The release history each exercise argues from is research/blog-era-translation.md: 198 rows of what the post says / what is true at the pin / changed in / the decision behind it. Exercises cite it rather than re-deriving eleven years of releases and each getting it slightly differently.

One warning that record earns, restated here because every exercise depends on it: Kubernetes' own CHANGELOGs and deprecation notices misstate their own removals, in both directions. The dockershim removal was announced for 1.22 and landed in 1.24. CHANGELOG-1.28 announces the azureFile removal that was reverted and actually happened in v1.30. kubectl run-container was announced dead in v1.8 and died in v1.13. So no exercise may source a removal release from the note that announced it — only from the source tree at the release tag. The announcement is not the event.

The census

Every post gets a row. All eleven of 2015's weekly-hangout notes, every launch party, every video roundup. A census that silently drops the noise is indistinguishable from one that missed it, and a reject costs one clause.

Rows live in YYYY/README.md, one file per year. The table shape is frozen so that twelve years of them concatenate without editing:

| post | date | k8s | verdict | topic | why |
|---|---|---|---|---|---|

post is the post's title, linked to its live URL, taken verbatim from manifest.tsv's url column and never assembled by hand — see the pin for why that column exists. date is MM-DD; the year is the file. k8s is the version the post targets, or — where it predates versioned releases. A row's link never changes once written, which is what lets a census pass emit its table and never revisit it; the exercise file is linked from the year's Exercises table instead, alongside its written/pending state.

An unpublished draft still gets a row. The eleven draft: true posts — 2026 only, the v1.37 queue — are post files at the pin, and every post gets a row does not carve out the ones the pin caught mid-flight. Their post cell carries the title unlinked followed by (draft), and the row is verdict-eligible like any other. draft: true is what decides that, not whether a URL resolved (ratified in #73): ten of the eleven are undated and resolve to no permalink, but the eleventh carries a date, so the template resolves a URL for it that the site returns 404 for. An unlinked row is the honest one. Ten of the eleven also carry no publication date, so they sort last — in the manifest and in the census — and the exercise numbering counts them last too. The reason to keep them is that the pin will move: when they publish, their rows are already written and the refresh is a link, not a census.

Verdicts:

verdict	means
walk	earns an exercise file
read	worth reading, nothing to do — a concept or an argument that outlived its commands, or a technique another exercise already carries
dated	was hands-on and would still be, but it does not fit the lab — hardware or scale beyond the 9.5GB ceiling
skip	no learning content — hangouts, launch parties, video roundups, recruiting, pure PR

dated is deliberately the narrowest verdict, and the only one that says the lab is why, not the post. It takes two tests, both of which must hold:

1. The post is itself a walkthrough — something a reader could follow end to end on bigger hardware.
2. What it needs is beyond the 9.5GB ceiling.

The second test alone is not enough, and this is the correction the 2024 census forced. Eleven of that year's read rows also fail on the lab — they want a CSI driver implementing ModifyVolume, or snapshots, or group snapshots; a real OIDC issuer; CRI-O instead of containerd; Windows nodes; NUMA or SMT topology; load a 4-vCPU guest cannot generate. None of them is a walkthrough. They are announcements with a snippet, so they are read: there is nothing to walk even on a bigger machine. Without test 1, dated swallows a third of every modern year and stops distinguishing anything.

A post whose technique is real but redundant, or whose tool is abandoned, is likewise a read — worth your eyes, not your cluster. Across the first two years censused, dated is 5 rows in 98. That is the right size for it.

Topic, from a fixed vocabulary of twelve: api sched net storage security nodes etcd obs tooling ecosystem history meta. obs is monitoring, logging, metrics and tracing — posts about seeing the cluster, which otherwise scatter across the subsystem each one happens to instrument.

Two years of real triage, 98 posts, settled three things about this list:

meta is the reject tag, and it is about a third of every year. Measured: 35 of 98. It is not one thing — it is hangout minutes, maintainer interviews, release announcements, and promo. It stays one tag anyway, because a topic tag exists to answer "show me everything about storage across eleven years", and that query is only ever run against walk and read rows. Splitting meta would buy three tags that appear almost exclusively on skip rows: cost, no query.

A topic earning zero rows in a year is expected, not evidence against the slot. obs matched four posts in 2015 and none in 2024 — the blog stopped covering observability as observability and scattered it into whichever subsystem is being instrumented. sched at 2 rows and etcd at 1 are not underused tags either; they are a 13% sample that happens to miss the years those subsystems got their attention. Twelve slots exist to make twelve years greppable, not to be evenly filled.

No tag went unused. Every one of the twelve earned at least one row across the two eras, which is the evidence that the list is neither too fine nor too coarse.

Four genres with a rule, because they are most of the corpus

Vendor and ecosystem posts — skip unless the post documents a decision and the alternatives it rejected; then read. Never walk. The test is answerable from the first screen, which is all the triage budget allows, and the bar on walk is the important half: an exercise that depends on a company staying in business is an exercise with an expiry date.

Weekly hangout notes — skip, as a class, all of them, in every year. Minutes carry a trace of an argument, not the argument; an exercise built on a trace has to supply the reasoning itself, at which point the note is a citation rather than a source.

Seeds are not lost by this: the row's why column names the seed, which is what makes it findable later. 2015's minutes record the kubectl exec demo entered via nsenter, and "treating rcs as pets vs. cattle" — the argument that produced the Deployment object — and both are written into their rows. Neither year has an exercise to footnote them in yet; when one is written, in any year, the note becomes its citation. A seed with no exercise to serve is still not a reason to promote the note.

Beta-graduation posts — the "Kubernetes 1.31: X Graduates to Beta" genre, 22 of 2024's 54 posts and the majority of every year from 2019 on. walk iff either:

• the feature's ladder has a discontinuity — a beta that shipped off by default, a withdrawn alpha, a permission that moved, a field that was renamed — or
• the failure the feature fixes is observable in the lab.

Otherwise read. A feature that marched alpha → beta → GA with nothing surprising in it has nothing an exercise can teach that the reference docs do not already say better.

The discontinuity clause is first because it is the clause that makes this archive worth having. A post that was right and stayed right teaches you its own year. A post that was right and then got quietly overtaken is the whole thesis — and the ladder is where you see it, which is why every exercise in this genre builds the ladder table the template specifies, out of the pinned docs' feature-gate frontmatter rather than out of the post.

Maintainer interviews — the Spotlight on SIG X genre, 9 of 2024's 54. read when the interviewee owns code, skip otherwise. The hangout-note rule above does not transfer: minutes are a trace because nobody was explaining, whereas an interview is a maintainer explaining a decision on the record, which is a source. "Owns code" is answerable from the title — a technical SIG on one side, a book club or a working group on the other. Never walk: an interview has no commands in it.

Release announcements — skip unless a specific line in the post is load-bearing for an exercise elsewhere; then read, and the why column names that line. This keeps the verdict earned against a fact instead of awarded to a genre — which matters most in the later years, where this genre is the majority of the corpus. Either way a release post is evidence of an announcement, never of an event: version numbers in an exercise come from the source tree at the release tag. See research/blog-era-translation.md for why that distinction is not pedantry.

Triage is title + frontmatter + first screen. The corpus is 7.22MB of post markdown, roughly 2.0M tokens — summed from manifest.tsv's bytes column, so it moves only when the pin does. It cannot be read in full by anyone, ever. Full reads are spent on walk candidates only.

A first-screen digest of a whole year costs between a sixth and a tenth of reading that year end-to-end. Measured across the whole corpus by digest.py — one instrument, one 900-character cut, so the years are comparable with each other:

year	posts	source	digest	cost of a digest
2015	44	292KB	47KB	16%
2016	90	681KB	100KB	15%
2017	53	460KB	58KB	13%
2018	70	714KB	76KB	11%
2019	52	551KB	57KB	10%
2020	56	587KB	61KB	10%
2021	53	582KB	58KB	10%
2022	69	638KB	75KB	12%
2023	78	709KB	85KB	12%
2024	54	568KB	59KB	10%
2025	78	777KB	85KB	11%
2026	70	832KB	76KB	9%

Whole corpus: 767 posts, 7.22MB of source, 837KB of digest, 11% — the same 7.22MB the paragraph above counts, because the script reads its source sizes from manifest.tsv rather than measuring them a second time. The estimate the sweep was budgeted against — 579 posts and 5.71MB left after the first three passes, digesting to 631KB at 11% — held to the end: every remaining year fitted one pass, and the last and cheapest, 2026, came in at 9%.

The ratio drifts down because the cut is a constant and the posts are not — a 2026 post averages 1.8× the length of a 2015 one, so a fixed first screen is a smaller slice of it. That is arithmetic about the instrument, not a finding about the blog. The finding about the blog is a different one, and it concerns triage quality rather than cost: the modern genre is fixed — what it was, what it is now, which gate, how to try it — so a later post's first screen predicts the rest of it far better than an early post's does. The census is still cut one year per pass rather than two, for the file-shape reason rather than the cost one — see the twelve years.

Known defects in the rubric

Five defects the twelve passes found in the rules above. All five are flagged and not applied, and that is a ruling rather than a backlog: each of them binds on verdicts, verdicts are immutable under #57, and the only sitting that could act on them is a re-pin, which #52 puts out of scope. They are written down because a reader who finds one of them for themselves should find it already named, not think the sweep missed it. Ruled in #78.

The first four are rubric questions — the rules would need rewriting before a verdict could change:

• ecosystem and tooling are not distinguished anywhere, and the boundary decides eligibility, because the vendor rule bars an ecosystem post from walk. Filing Helm as a project's tool rather than a company's product is what let Helm Charts take an exercise. The case that shows the seam is Stern, whose company died and whose tool outlived it. (#64)
• The rules are ordered and nothing says in what order. dated requires a walkthrough, and the vendor rule reaches its candidates first; applied in that order it cost dated its only row in both 2017 and 2021. Reverse the two and those years read differently. (#65, #69)
• The release-announcement rule's "load-bearing line" escape hatch cannot fire on a version fact, because #54 forbids sourcing a version from an announcement. In practice the genre has collapsed to a class-wide skip, and the rule should say so rather than describe an exception that cannot be taken. (#65)
• The maintainer-interview rule tests owns code, where the test actually applied was process versus engineering. SIG Node's spotlight interviewees own the kubelet and the post is pure process, so the written rule and the applied rule disagree on the exact case the rule exists for. (#66, #69, #73)

The fifth is not a rubric question. It is a reader-visible defect in a shipped artifact:

• The spotlight rule was applied inconsistently across five years. 2024 and 2026 follow what is written; 2020, 2021, 2022, 2023 and 2025 skipped every spotlight they met. So two spotlight posts of comparable substance can carry different verdicts for no reason a reader can recover from the rules, and the census file is where they will meet it. Recorded here rather than repaired, on the same immutability rule as the four above. (#73)

The budget

8–15 walk verdicts per year — roughly 100–180 exercises against labs/'s 357. No hard cap. Each year's README states what it spent, and a year outside the band argues for itself in one line.

It was a budget, not a yield, and all twelve years have now spent it: 135 walk across 767 rows. Both ends were reached — the floor of 8 in 2017 and 2019, the ceiling of 15 in 2016 — and in twelve passes the band never once cut a post the rubric would otherwise have kept. Every low year was low because its posts were release notes; every high year was high because its posts were walkthroughs. So the band described the corpus for twelve years and never disciplined it. The walk set is closed, which means it now binds on nothing: what follows is a premise the authoring passes inherit, not an open question.

The band is deliberately not re-cut per era or per release. Its whole function is friction against a rubric that will drift generous over twelve passes, and a band that flexes with the year's density stops being friction.

The twelve years

Twelve census passes, one per year, all closed. Each row links that year's full census; the walks column is what the year owes in exercise files, and the last is what the year is — which is how to choose an entrance if chronological is not what you want.

year	posts	walks	what the year is
2015	44	10	Nothing runs as written. No RBAC, no Deployment, no kubeadm, no CRI, and kubectl apply does not exist yet
2016	90	15	The largest year, and the one whose nouns all still exist meaning something else — PetSet arrives and is renamed StatefulSet nine months later
2017	53	8	The security defaults move: RBAC beta in 1.6 and GA in 1.8, NetworkPolicy stable in 1.7, containerd a real option under CRI
2018	70	9	The storage stack rebuilt in public — CSI runs alpha, beta and GA inside twelve months
2019	52	8	Where the archive starts going wrong in a new way: posts that were broken on the day they were published
2020	56	9	Where it starts arguing with itself — a post promising an API "is not going away" that the pin marks deprecated
2021	53	11	The blog reorganises around the release train. The genre changes here, and this is the second entrance
2022	69	13	The year that proves the ladder is not one-way — three of its alphas are replaced rather than promoted, and their gate ladders restart
2023	78	13	The most evenly spread year in the archive, and the one that found a second feature-gate registry inside kubeadm's own docs
2024	54	13	Almost everything still runs, which is the opposite problem — the diff is a gate's graduation, not a broken command
2025	78	12	The first year whose largest topic is the software rather than the project; 63% of its titles name a version
2026	70	14	Two-thirds of a year — the pin cuts it at 2026-08-26 — and eleven of its rows are unpublished drafts

A pass is one year because a census file is one year. The map originally cut these as five two-year eras of 109–148 posts, and the prototypes proved that wrong by censusing 44 and 54: an era is 2.0–2.7× anything anyone had run, and a pass that runs out of room at post 110 of 147 leaves a half-filled YYYY/README.md in a state no other pass can pick up. Cutting by post count instead was rejected for the same reason — it puts a file boundary inside a pass. Whoever refreshes the pin inherits that constraint, and a thirteenth year with it.

Authoring progress is not recorded here. It lives in each year's ## Exercises table, one written or pending per row, which is the copy the gate checks against the files on disk in both directions — so the gate's own run is where you read how far along the tree is. A second table in this file would be a copy nothing checks.

The one link out

Exercises cite strands/lab-topologies.md for topologies and for provision and teardown steps, rather than restating them. That is the only coupling in either direction. Duplicating the 9.5GB ceiling into a second tree is how you end up with two contradictory answers to "how much RAM do I have".

strands/check-anchors.py is not extended to this tree. blogwalk/check-census.py is this sweep's only gate, and it checks census rows against manifest.tsv.

The gate

python3 blogwalk/check-census.py

Run it from the repo root before closing a census pass. It is the sibling of strands/check-anchors.py — same invocation style, same exit codes, green in both directions — and it exists because at 767 rows, swept and feels swept are indistinguishable by eye.

manifest.tsv is the authority. Every row is matched back to a manifest row, so a post nobody censused and a row for a post that does not exist are equally loud, and the match is on the url column byte-for-byte — a case-only difference in one of the eleven literal permalinks fails. It also reads the four verdicts, the twelve topics and the budget band out of this file, so the rubric lives here and not in a second copy inside the checker; if it cannot parse them it fails rather than checking nothing.

It reports authoring progress and does not judge it: the per-year table carries written and pending beside walk, counted from each year's ## Exercises table — the same cells classes 3 and 4 already cross-check against the files on disk — and a closing line states how many of the 135 exercises exist. Nothing here fails because an exercise is unwritten, on the same reasoning as the pending year below; but the map's authoring condition, every walk verdict has its exercise file, is visible from nowhere else, so the gate says it outright rather than leaving the last authoring pass to work out that it was the last one.

An uncensused year is pending, not broken. The gate was green with ten years still to sweep, which is what made it usable as a gate while the sweep ran, and it is green now with none left. Read the per-year table it prints: posts comes from the manifest, rows from the census, and a year whose walk count leaves the budget is flagged on the run rather than on inspection.

The one judgement it cannot make is dated. That verdict takes two tests and only the second — beyond the 9.5GB ceiling — is mechanically visible at all, so a dated row whose why cites no hardware and no scale is reported as a warning: it asks a human to look, and does not decide.

Layout

blogwalk/README.md          this file — method
blogwalk/TEMPLATE.md        the exercise template, forked from labs/
blogwalk/manifest.tsv       one row per post at the pin
blogwalk/build-manifest.py  regenerates the manifest, asserts the counts
blogwalk/check-census.py    the gate — run it before closing any census pass
blogwalk/digest.py          builds a year's triage digest; the reading budget, made mechanical
blogwalk/YYYY/README.md     that year's full census
blogwalk/YYYY/NN-slug.md    one exercise per `walk` post, numbered in publication order