Method details

adaptiSeq fetches every sequence-data byte through a single download seam, so the scheduler and engines change only how bytes arrive and when files are scheduled — never which bytes. Resolution, metadata, integrity, logs, and merge are independent of the transport in use.

Engines

Engine	Flag	What it is
segmented (default)	--engine segmented	Resumable, range-based HTTP(S)/FTP downloader (aiohttp/aioftp). Each file is fetched in multiple byte-range connections with atomic .part/.part.meta resume. Runs the batch worker pool and adaptive controller.
classic (opt-in)	--engine classic	The plain-tool path: wget, or axel with -p, plus ascp for -a. No segmentation.

The segmented engine never falls back to classic. A host that cannot serve ranges is downloaded as a single stream inside the segmented engine; --engine classic is a manual choice only.

Transport selection (segmented, protocol auto)

For ENA the default is HTTPS-first:

1. Prefer the HTTPS mirror, confirmed by a cheap per-host range probe.
2. Else native segmented FTP (REST/RETR via aioftp).
3. Else a single stream inside the segmented engine.

-r https / -r ftp overrides and is final. A corrupt or zero-byte file is never produced. Per-host transport is cached by kind, not URL (a deliberate fix — a URL cache made every file on a host download the first file's bytes).

Download methods summary

Method	Used for	Notes
segmented-https	ENA/most HTTPS mirrors	default; multi-range, resumable
segmented-ftp	ENA FTP (REST/RETR)	when HTTPS ranges unavailable
single-stream	range-incapable hosts	inside the segmented engine; never truncated
aspera (ena)	ENA, -a	adaptive ascp pool (efficiency hysteresis)
aspera (gsa)	GSA, -a	sequential, Huawei-Cloud preferred; best-effort
huawei-cloud	GSA	preferred for GSA even with -a
classic wget/axel	--engine classic	the plain-tool path; axel with -p

Concurrency & etiquette

• Per-file: up to --max-segments byte-range connections.
• Per-host: a global cap (--max-conns-per-host) plus a reactive circuit breaker (429/503/refused → exponential global backoff + temporarily lowered cap, slow recovery).
• Per-batch: a worker pool (-j) whose active size the adaptive controller tunes (--adaptive; topdown knee search by default, bottomup gradient via ASEQ_ADAPTIVE_MODE); workers are gated at file-pickup boundaries.
• Resolution: parallel (--meta-jobs) under per-endpoint rate limiters (ENA / NCBI / GSA; NCBI 3 rps, 10 with NCBI_API_KEY).
• Speed: -s/--speed token-bucket cap in MB/s.

Integrity

Every Run is md5-checked against the public database (SRA uses md5 + vdb-validate). On mismatch adaptiSeq retries up to three rounds, then records the outcome in success.log / fail.log. -k/--skip-md5 disables the checks.