Fire in many caves — all light, or none light

Ugg teach big magic today: move shiny stones between many caves, and never lose one stone.

A distributed transaction — one job that touch stones in many caves at once. Ugg move 100 shiny stones from Cave-A to Cave-B: take from A, add to B. Must BOTH happen, or NEITHER. Only take happen — 100 stones vanish! Only add happen — stones double! Either way, magic broken.

The cave where job start, that cave is the coordinator — the chief. Every other cave holding stones is a participant — a hunter in the job. Chief boss the job; hunters do their part. To force all-or-nothing across caves, tribe use a commit protocol — every cave must vote to commit BEFORE any cave really commit.

Many caves mean new ways to break: a hunter fall down (site failure), a grunt get lost in wind (lost message), the path between caves cave-in (link failure), or a rockslide split tribe into two groups who cannot hear each other (network partition). Nasty part: to each side, the other side look just like it fell down.

ACID — the four tribe laws

Every stone-job must obey ACID:

$\text{ACID}=\text{Atomicity}+\text{Consistency}+\text{Isolation}+\text{Durability}$

Atomicity — all-or-nothing, whole hunt or no hunt. Consistency — never break tribe rules; stones counted same before and after. Isolation — two jobs at once no peek at each other half-done mess. Durability — once job done, done forever; even fire in cave no undo it. Across many caves, Atomicity is the HARD law — that is why 2PC and 3PC exist.

Two-Phase Commit — chief count grunts

Two-Phase Commit (2PC) is old reliable way. It trust the fail-stop model: broke caveman just fall and stop — he never lie, never send wrong grunt — maybe wake later. Every cave carve marks on a stone tablet (stable log) that survive fire.

Phase 1 — Prepare / Voting. Chief carve $⟨\text{prepare }T⟩$ on tablet, force to stone, then howl PREPARE to all hunters. Each hunter think: can I do my part? If yes — carve $⟨\text{ready }T⟩$ to stone, grunt READY back. If no — carve $⟨\text{no }T⟩$, grunt ABORT.

Phase 2 — Decide. If chief hear READY from ALL hunters — carve $⟨\text{commit }T⟩$ to stone, howl COMMIT to all. Else — carve $⟨\text{abort }T⟩$, howl ABORT. Every hunter carve the word and do it.

The moment $⟨\text{commit }T⟩$ touch chief's tablet is the point of no return:

$⟨\text{commit }T⟩\text{on chief’s stable log}=\text{decision forever}$

Once chief carve COMMIT in stone, no fire, no falling-down, no anything take it back.

Reading the tablet after you wake

Hunter fall mid-job, wake, look at his tablet:

• See $⟨\text{commit }T⟩$ → redo(T), do the job.
• See $⟨\text{abort }T⟩$ → undo(T), take it back.
• See $⟨\text{ready }T⟩$ only, no decision → ask chief. This is the scary in-doubt state.
• See nothing → he never voted, chief must have aborted → undo(T).

In-doubt is bad. Hunter voted READY, so he still GRIP all the meat and berries (locks) the job touched — no other job may eat them — until he learn the word. Tablet keeps the lock list as $⟨\text{ready }T,L⟩$ so he grab them all back when he wake.

The blocking problem — hunters frozen

Now worst fire. All hunters carved $⟨\text{ready }T⟩$. None got the decision. And the CHIEF FELL DOWN.

$\forall \text{awake hunter: tablet}=\{⟨\text{ready }T⟩\}\wedge \text{chief crashed}\Rightarrow \text{BLOCK}$

Why frozen? Chief maybe carved $⟨\text{commit }T⟩$ before falling — then some far hunter maybe already committed, so ALL must commit. OR chief fell before carving — safe to abort. Awake hunters cannot tell which! None can guess safe. So all WAIT — frozen, still gripping meat. That is the blocking problem, the big weakness of 2PC. Same trouble with a rockslide (partition): cut-off caves block, though never a wrong answer — chief just treat missing caves as fell-down and run on his side.

Ugg warn: 2PC also cost extra stone-carving (slow), and chief is single point of break.

Three-Phase Commit — tell more caves first

Three-Phase Commit (3PC) add one step so hunters can decide WITHOUT chief. It need strong luck: no rockslide (no partition), always at least one cave awake, at most $K$ caves break.

Phase 1 (PREPARE) — same as 2PC. Phase 2 (PRE-COMMIT) — chief hear all READY, howl PRE-COMMIT, and WAIT for at least $K$ hunters to grunt back before going on. Now the "we gonna commit" word carved at $K+1$ caves. Phase 3 (COMMIT) — chief howl final COMMIT (or ABORT); hunters do it.

If chief fall after Phase 2, awake hunters check each other:

$\exists \text{awake cave with}⟨\text{pre-commit }T⟩\Rightarrow \text{commit};\text{else abort}$

Someone hold $⟨\text{pre-commit }T⟩$? Pick new chief, howl COMMIT — no freeze. Nobody hold it? Word never went out, safe to abort. **3PC no block as long as $\le K$ caves break.**

Why tribe still use 2PC

3PC cost extra round of howling — more grunts, more carving, slower. Worse, it ASSUME no rockslide — but rockslides happen! In a partition, both sides may decide different — magic broken. So real tribes run 2PC plus strong chief-picking magic (Paxos or Raft): caves share the chief job, and when chief fall they vote a new chief fast. 2PC underneath, no single point of break, no need for 3PC's silly luck.

Ugg warn — no do this!

• No say "2PC handle rockslide." It NO — partition make hunters block.
• No say "3PC is what everyone use." It NOT — no-partition luck too much; tribes use 2PC + Raft/Paxos.
• No mix up $⟨\text{ready }T⟩$ ("I vote yes," held by hunter) with $⟨\text{commit }T⟩$ (chief's final word). Different stones!
• No forget: both $⟨\text{ready }T⟩$ and $⟨\text{commit }T⟩$ must be FORCED to stone BEFORE you grunt — else waking hunter lose the memory.

Ugg remember

• Atomicity across caves is the whole point: commit at ALL caves or abort at ALL caves, never half.
• 2PC = PREPARE + DECIDE. $⟨\text{commit }T⟩$ on chief's stone = point of no return.
• Wake-up tablet rule: $⟨\text{commit}⟩$ redo / $⟨\text{abort}⟩$ undo / $⟨\text{ready}⟩$-only ask chief / nothing abort.
• 2PC blocks when all hunters $⟨\text{ready}⟩$ AND chief crashed — frozen, gripping locks.
• **3PC = PREPARE + PRE-COMMIT ($K$ acks) + COMMIT.** Pre-commit at $K+1$ caves breaks the block — but assumes no partition, so real tribes run 2PC + Raft/Paxos instead.