One Pilgrim, One Plate
When Maharaj Pakacharya first ran the temple kitchen, he served pilgrims one at a time. A pilgrim arrived; the Maharaj lit the tandoor; he kneaded dough; he cooked the meal; he served it; he cleaned the tandoor; he waited for the next pilgrim. The food was magnificent. The line outside the kitchen wound around the temple twice. By dusk, three hundred pilgrims had eaten and seven hundred had given up.
“Excellence served sequentially is famine for everyone but the first.”
The Stale Batch of the Old Cook
A travelling cook offered a different way. "Cook a great vat of dal at dawn," he said, "and serve it through the day to whoever comes." The Maharaj tried it. The line moved swiftly. But by noon the dal had cooled, and by dusk it had soured. The pilgrims who came late ate what was no longer good. "This," said the Maharaj, "is worse than the line."
“A batch that grows stale before service is worse than no batch at all.”
The Continuous Tandoor
Then came a wise potter's daughter who had once worked in a wedding kitchen. "Your tandoor has many sides," she said. "While one chapati is rising on the inner wall, slap the next on a free spot. As one comes off, another goes in. The tandoor is never empty, but it is also never crowded. Each chapati cooks at its own pace; do not wait for a whole batch to finish before starting the next." The Maharaj tried it. The line vanished. Each pilgrim received a hot, fresh chapati within minutes. The tandoor was hot all day, but never wasted.
“A tandoor served continuously, with each loaf added as another finishes, feeds many without making any wait.”
The Shared Spice Box
But the Maharaj noticed waste in another corner. Each cook had his own spice box, and each opened the same jar of cumin a hundred times a day. The potter's daughter again: "Keep one shared spice tray on a turning stand at the centre, with each spice in its own bowl. Any cook may reach in for any spice. No cook hoards a private jar." The kitchen quickened further. The same spices, used by many, never wasted, never duplicated.
“A pool of memory shared among many requests serves more than the same memory copied into each.”
The Speculative Apprentice
A young apprentice was sent to help. He could roll chapatis quickly but could not judge their finish. The Maharaj used him cleverly: "Roll out three chapatis ahead of what I have asked for. If I want them, they are ready. If I do not, throw them to the cows; they will not be wasted." The kitchen sped up further still — most of the apprentice's rolls were used; the few that were not, the cows enjoyed. The pilgrims felt their food arrive almost before they had asked for it.
“A cheap apprentice that prepares ahead and is sometimes wrong is faster than a master who prepares only on demand.”
The Kitchen That Did Not Crash
On the day of the great festival, ten thousand pilgrims came. The kitchen, with its continuous tandoor, its shared spice tray, its speculative apprentice, fed them all by sunset. The Maharaj sat afterwards with the potter's daughter on the temple steps and said, "I learned three things. The tandoor must never be wasted but never crowded. The spices must be shared, not hoarded. And the apprentice must be allowed to guess. None of these I would have invented alone. None of these the old way of one-pilgrim-one-plate would have allowed."
🪔 Deepak — the lamp of meaning · what this fable means in code
The Maharaj's one-at-a-time kitchen is a model server with no batching — perfect quality, terrible throughput. The travelling cook's stale batch is naive static batching that waits for a fixed group, raising p99 latency for early arrivals. The continuous tandoor is continuous (in-flight) batching, the breakthrough of vLLM and friends — new requests slot into a running batch as old ones finish, keeping the GPU full without ever making a request wait for the next batch boundary. The shared spice tray is PagedAttention and prompt-prefix caching, where memory for shared context is pooled rather than duplicated per request. The speculative apprentice is speculative decoding — a cheap draft model proposes several tokens ahead, the main model verifies in parallel, and most drafts are accepted. Together these turn a kitchen that could feed three hundred into one that feeds ten thousand without losing the heat of any single chapati.

