Interleave serverless requests across collections

[generall]

What

serverless upload and serverless query drained collections sequentially — the whole budget of collection 0, then collection 1, and so on. With bounded parallelism only one or two collections ever saw concurrent traffic, which isn't the multi-collection workload the mode is meant to produce.

• New drain_budgets in serverless/distribution.rs: pre-allocate per-collection budgets as before, then emit requests in random order among collections that still have budget. Each collection keeps one contiguous id range, so a dataset source still gives every collection a distinct slice.
• Sampling while draining is uniform on purpose — the uniform/Zipf skew already lives in the budgets; re-applying the picker there would skew twice.
• query used to pick a collection per request, so actual per-collection query counts were a random walk rather than the requested distribution. It now walks the same pre-planned batches.
• Requests are no longer all batch_size items, so Processor gains request_count/request_size (defaults = the old arithmetic) and the progress bar asks the processor instead of assuming a fixed size.
• plan_batches in upload.rs is replaced by drain_budgets; its test moved over.

Checks

• Targets dev
• cargo test --all — 176 passed
• cargo +nightly fmt --all
• cargo clippy --all --all-features — clean
• New test draining_budgets_preserves_limits_offsets_and_totals covers per-collection totals, batch cap, and contiguous offsets

🤖 Generated with Claude Code