Labs — Phase 2, etcd internals

Twenty-nine exercises in the order they are meant to run. Each states one claim to test or one artifact to produce, links its topology rather than restating a footprint, and ends with a teardown that does two things: deletes what that exercise created, then says whether the topology stays or goes. Here it almost always stays — one cluster runs the whole phase.

The framing — why each module exists, what to read and the question to answer from it — stays in phases/02-etcd.md. These files hold only what you type and what you should see.

#	Exercise	Why it exists
1	Two clones, and an entry point that resolves	The tree every citation for the next month is true against, plus the sha that makes it checkable.
2	The release that moved mvcc/ under server/	git log --follow across a rename, cross-checked from the other end so one technique is not one result.
3	Raft is a second repository	--diff-filter=D finds the commit that deleted a path — the only commit the current tree will never show you.
4	Break one published citation, on purpose	The pickaxe, and the first time the required answer is a sha rather than a paragraph.
5	Three members, wired together by hand	The flags the restore drill turns on, met three weeks early as "the flags that made it start".
6	Predict which historical reads survive a compaction	Gate condition 3. The prediction is the exercise; the run only marks it.
7	One raw bbolt key, decoded	Where the bbolt key turns out to be the revision and the value carries the user key — the reverse of the usual assumption.
8	The transaction that puts two revisions inside one	sub is invisible to the API, which is why the previous exercise had to reach for raw bytes.
9	The index is in memory, and it is not free	Deleting keys does not return the memory. Compaction is the only thing that does.
10	A stale read you caused on purpose	Closes the question P1 deferred — and shows a wrong answer arriving with no error attached.
11	One leader, three logs, and a handover you asked for	The gap between committed and applied, which every later module lives in.
12	An election you did not ask for	A failed write is not a write that did not happen — consistent is not exactly-once.
13	What the library obliges its caller to do, in order	No disk, no network, no clock. The library counts ticks, and that is why slow clusters elect.
14	The leader write path, with the fsync in the right place	Why a leader may send before it syncs and a follower may not — the sharpest thing in module 2.2.
15	2.C5 — cut one member off the peer port	A member that is healthy, answering, and refusing to serve. The client-side network is the one segment that works.
16	A watch that starts in the past	Nothing marks the boundary between replay and live — which is why informers list before they watch.
17	Three sets a watcher can be in	Catching up is slower per event than being live, which is the opposite of most intuitions.
18	Delete the history a watcher is still standing in	Two callers, two consequences, one error value — and the split this module was written to force.
19	One compaction, four layers, and the relist at the top	Gate condition 1. The only exercise in the curriculum graded on recall.
20	Compaction moves one number and not the other	"We compacted and the disk is still full" is not a bug in any layer.
21	Defrag returns the disk and stops the member	One writer at a time means there is no incremental version of it. Never --cluster.
22	History that disappears while you are not looking	The configuration a real cluster ships with, and the reason ErrCompacted is usually nobody's fault.
23	2.C4 — drive it into NOSPACE and get it back	A delete is a write, so a delete is refused. The recovery has exactly one correct order.
24	Three kinds of file, and which one you cannot lose	A member that cannot be repaired is removed and re-added, not fixed.
25	Predict which timing this hardware breaks	"etcd is slow" is not a claim until it names a histogram. The 35 W box fails the way production does.
26	2.C1 — write garbage into one page	The quiet corruption is the dangerous one: healthy, elected, and serving different answers.
27	2.C2 — stop two of three	--force-new-cluster is not a repair; it is an operator stepping outside the protocol.
28	2.C3 — put the cluster back, and count what you lost	A restore is a rollback of the whole keyspace, with no error naming what went.
29	The capstone — one sitting, two artifacts, four citations	Graded as much on whether the file:line references survive checking as on whether the cluster came back.

One cluster runs exercises 5 through 29. etcd-only comes up at exercise 5 and is destroyed at exercise 29, which ends the phase. It is 3.0GB of a 9.5GB budget and there is no Kubernetes anywhere in this phase — no kubelet, no CNI, no controllers. That is the whole reason the plan calls this the best depth-per-MB month in the curriculum, and it is why a phase this long fits a 35 W homelab at all.

Exercises 1–4 and 13 need no topology. They run on forge, which is up regardless, and the two blobless clones they create live there for the whole phase and outlive the cluster — every file:line in the capstone is a claim about a sha in one of them.

The cluster is deliberately unencrypted, which is a stated deviation with its reasoning in exercise 5: everything here is one --cacert/--cert/--key triple short of the command you would run against a kubeadm cluster's etcd.

Three exercises leave the store deliberately dirty for the next one — 17 keeps its keys for 20, and 20 keeps its inflated db for 21, because a compaction with nothing to compact measures nothing. Their teardown lines say so; do not tidy past them.