H2HDB Downloader (h2hdb-downloader)

Automates downloading galleries from exhentai/e-hentai (via hbrowser>=0.44.0,<0.45.0) and recording their state in an h2hdb database. It has no CLI or standalone runtime of its own — it's a library consumed by another project that owns the browser session and the overall process lifecycle.

Concepts

• Gallery — a single exhentai/e-hentai gallery, identified by a gid (numeric id) and represented as h2h_galleryinfo_parser.GalleryURLParser once its URL is known.
• Dedup — before issuing a real network download, the package reads live h2hdb state to see if the gid is already settled (downloaded, with no redownload flag or durable request). Settled gids are skipped — except periodically, at a random interval (1 to 19 attempts), when one is force-redownloaded as an integrity re-check.
• Durable requests — immediately before a real download starts, the package creates a tokenized request through h2hdb's normalized operational queue facade. It conditionally completes that exact token only after success. A False result, exception, cancellation, or process termination leaves resumable work behind, while a newer request for the same gid cannot be erased by an older attempt finishing late. h2hdb also uses this table to publish a redownload request after all active deletion-candidate folders for a gid have actually disappeared. For a deep root job, success means that the root gallery has resolved and its entire related-tag cascade has returned successfully; the root request remains queued throughout that traversal. Before drain_pending_redownloads() performs any browser or network work, it walks its entire pending snapshot and calls ensure_download_request() for every GID, turning each item into a durable tokenized root request. If the process stops during this seeding pass, already seeded roots are recoverable through the durable queue and unseeded roots remain in the pending-redownload view. Pre-seeding also means an earlier root's cascade reuses, rather than completes, the token belonging to a later pending root; that later root still runs its own cascade before its token is settled. A public single-root call performs exact-token deletion and the download-to-ingest handoff in one transaction. A drain_queue() batch instead checkpoints each returned root through its still-live turn, then performs one handoff at the batch boundary. A late worker that has lost its turn therefore cannot delete the recovered root request. If another caller has replaced the request token while the turn is still valid, that newer request remains queued. A GID is recorded as removed only from hbrowser's explicit ConfirmedGalleryMissing result, never by interpreting an empty or malformed page. Missing-marker writes and exact-token deletion are atomic; the single-root form includes the handoff in that transaction, while the batch form retains DOWNLOADING until the boundary. A newer request fences both missing mutations.
• Core boundary — the caller injects h2hdb's public VNextDownloadQueueFacade from h2hdb>=0.29.0,<0.30.0. This package never opens a connector, reaches into a repository, migrates the schema, or manages the database gate. Browser search, downloads, retry sleeps, and tag traversal remain outside the coordinator's short synchronous calls.
• Ingest backpressure — each public deep-download root still owns one h2hdb download turn, while drain_queue() and drain_pending_redownloads() group complete roots until at least download_submissions_per_ingest unique H@H submissions have been accepted. This is a soft threshold checked only after an indivisible root and its entire related-tag cascade return: if consecutive roots submit 10, 11, and 103 galleries with a threshold of 100, all three finish and the batch hands off with 124 submissions. A root that produces no accepted submission does not advance the threshold. A heartbeat spans the entire root or batch. Between batch roots, successful and confirmed-missing dispositions are persisted through the live turn fence without releasing DOWNLOADING; unresolved roots stay queued. The batch also hands off at snapshot exhaustion. Failures and cancellation attempt one immediate handoff without removing the interrupted root. If the process or container is killed before it can do so, lease expiry lets h2hdb recover and scan already published files. Each completed root is a durable checkpoint, so restarting discards only the process-local submission count, not correctness. Related downloads atomically reuse an existing request token instead of replacing a later snapshot root, and one drain snapshot remembers accepted GIDs so overlapping cascades neither resubmit nor recount them before ingest. This coordination uses only short core calls; browser work never holds a database transaction. Backend-neutral temporary unavailability is retried only at the ready-turn claim and completed-generation polling boundaries; backend lock handling remains inside h2hdb core.
• External submission semantics — H@H is treated as an uncontrollable external downloader. If it accepts a submission and this process stops before the corresponding database checkpoint commits, the next run may submit that GID again. Queue and ingest state remain correct, but H@H submission is deliberately at-least-once rather than exactly-once.
• Manual queue — add a (gid, url) row to the CSV configured by csv_path. It is converted into the same durable request and picked up the next time the queue is drained. Before replay, the inbox is atomically rotated to a same-directory hidden claim file; interrupted claims are replayed automatically on the next run. A blank GID is derived from the URL with GalleryURLParser before the request reaches core. When both fields are present, the URL must identify the same GID; malformed or mismatched rows are rejected before their claim is acknowledged.
• Deep download — download a gallery, then look at its artist/group tags and download sibling galleries that match a set of search conditions (e.g. other-language releases of the same work).

API

Downloader is the public service object. TagCascadePolicy and DownloadTurnLostError are the other public exports. Every method either acts on a target you explicitly pass in or, for the two queue-reading methods below, hands back a plain value with no further bookkeeping required from you. There is no "run the whole thing" method: deciding when to stop, what order to process things in, and how to report progress is the calling application's job, not the library's.

Downloader(
    driver: ExHDriver,         # an un-entered driver; see below
    facade: VNextDownloadQueueFacade, # initialized public h2hdb facade
    csv_path: str | None = None,  # path to the manual download-queue CSV
    *,
    wait4client: int,       # seconds to wait before retrying after ClientOfflineException
    retry2download: int,    # seconds to wait before retrying after InsufficientFundsException
    turn_poll_seconds: float = 5,       # wait interval for a turn / ingest completion
    turn_lease_seconds: int = 300,      # recoverable ownership lease
    turn_heartbeat_seconds: float = 60, # renewal interval; shorter than the lease
    download_submissions_per_ingest: int = 100, # accepted unique submissions
)

The application owns core configuration and startup. Inject an h2hdb>=0.29.0,<0.30.0 VNextDownloadQueueFacade connected to an epoch-3/schema-version-2 database; downloader never initializes the schema or loads core configuration. Older core compatibility lanes are not supported.

csv_path only enables the optional "queue a gid/url by editing a CSV file" feature described above. Leave it as None if you don't need that; durable database requests and live deduplication still work.

The turn timing defaults normally need no adjustment. All three timing values must be positive and finite, turn_lease_seconds must be an integer, and the heartbeat interval must be shorter than the lease. download_submissions_per_ingest must be a positive integer. It counts unique GIDs for which driver.download() returned True during the current batch; it does not measure H@H processing time or materialized gallery folders. The threshold is soft because a root and its cascade are never split. Roots with no accepted submission do not advance it.

Coordinated methods raise DownloadTurnLostError if their lease can no longer be renewed or the conditional handoff proves that another process owns the turn. Callers may catch it separately from browser/download failures; the durable root request for unfinished work remains available for a later retry.

Downloader is itself an async context manager that opens and closes the browser session for you, so driver is expected un-entered:

async with Downloader(ExHDriver(headless=False), ...) as downloader:
    ...

If you'd rather manage the driver's lifecycle yourself, pass an already-entered driver and skip async with downloader.

Method names follow one rule throughout: no suffix means it operates directly on a GalleryURLParser you already have; _by_gid means it resolves a bare gid through hbrowser's exact typed lookup first, then does the same thing.

• await download_by_gallery(target) — download one GalleryURLParser, or an iterable of them. Returns {gid: downloaded} for each. Retries automatically on ClientOfflineException (waits wait4client seconds) and InsufficientFundsException (waits retry2download seconds); a wait of 0 means "don't retry, raise immediately." This is a direct API and does not claim a download turn or wait for h2hdb ingest.
• await download_by_gid(gid) — resolve a bare gid through hbrowser's exact lookup, then download it. Only an explicit, independently confirmed missing result is recorded as removed in h2hdb; challenge, authentication, malformed, pagination, navigation, and bounded-search failures raise and leave the request retryable. A later successful lookup clears any stale removed marker. If the gid resolves to a different gid (the gallery was merged/redirected), the original gid is flagged for deletion after the replacement downloads successfully. This is also a direct, uncoordinated API.
• await download_by_tag(tag, conditions) — download every gallery under a hbrowser Tag, once per search condition in conditions (or unconditionally if conditions is empty). This is also a direct, uncoordinated API.
• await deep_download_by_gallery(gallery, policy, skip_check=False) — download gallery, then for each tag in policy.filters (e.g. "artist", "group") on that gallery, call download_by_tag with policy.conditions. The cascade only runs if the initial download actually happened, unless skip_check=True forces it to run regardless (useful when you already know the gallery is downloaded from a separate call and just want the cascade). policy is a TagCascadePolicy(filters, conditions) — both fields always travel together, so they're grouped into one frozen value object rather than two parallel parameters. The whole call is one coordinated root: it claims a turn, keeps its durable root request until the cascade finishes, atomically finishes the exact request while handing the turn to h2hdb, and waits for that generation to be ingested.
• await deep_download_by_gid(gid, policy, skip_check=False) — same gid-resolution as download_by_gid, but deep and coordinated as one root.
• await drain_queue(policy, skip_check=True) — absorb the manual CSV and process one live snapshot of durable database requests. A request is removed only after a successful root and complete related-tag cascade, confirmed removal, or successful redirect. Complete roots share one download turn, heartbeat, handoff, and h2hdb ingest wait until their accepted unique H@H submissions reach the soft download_submissions_per_ingest threshold. A URL-to-gid fallback remains part of its root traversal. Stale snapshot tokens are skipped without claiming a turn, and the method does not loop for newly queued work after its snapshot. If a malformed or mismatched URL is encountered despite the write-boundary validation, it is ignored and the original requested GID is resolved safely.
• await drain_pending_redownloads(policy, skip_check=True) — process one live snapshot of GIDs h2hdb flags for periodic redownload. It first seeds every snapshot GID as a durable tokenized request, before any browser/network work, then uses the same submission-count batching, indivisible-root semantics, durable checkpoints, and final ingest barrier as drain_queue(). An interrupted seed pass leaves seeded GIDs in the durable queue and unseeded GIDs pending. Newly flagged GIDs wait for the next call.
• pending_redownload_gids() — a snapshot list of gids h2hdb currently flags as needing a periodic redownload. Every call reads live database state through bounded, revision-and-cutoff-pinned keyset pages; read-only and safe to call repeatedly. Empty intermediate pages do not truncate the scan. Prefer drain_pending_redownloads() when processing the whole snapshot so it can share ingest barriers across roots.

Example

The calling application owns the loop. A typical one drains the durable queue once, then drains one pending-redownload snapshot. Both operations apply the same accepted-submission soft threshold:

import asyncio
from h2hdb_downloader import Downloader, TagCascadePolicy
from h2hdb import VNextDownloadQueueFacade, load_config
from hbrowser import ExHDriver
from h2h_galleryinfo_parser import GalleryURLParser

policy = TagCascadePolicy(
    filters=("artist", "group"),
    conditions=("language:chinese$", "language:speechless$"),
)


async def main():
    facade = VNextDownloadQueueFacade(load_config("h2hdb-config.json"))
    async with Downloader(
        ExHDriver(headless=True),
        facade=facade,
        csv_path="todownload_gids.csv",
        wait4client=30 * 60,
        retry2download=4 * 60 * 60,
        download_submissions_per_ingest=100,
    ) as downloader:
        gallery = GalleryURLParser("https://exhentai.org/g/123/456/")
        await downloader.download_by_gallery(gallery)
        await downloader.download_by_gid(666)
        await downloader.deep_download_by_gallery(gallery, policy)

        await downloader.drain_queue(policy, skip_check=True)
        await downloader.drain_pending_redownloads(policy, skip_check=True)


asyncio.run(main())

License

This project is distributed under the terms of the GNU General Public License version 3 (GPLv3). See the included LICENSE file for the complete terms.