dp_update() no longer treats an untracked legacy
copy as current merely because its local timestamp is later than the
registry date. A missing, unreadable or unverifiable identity record now
requires a fresh download when update is requested. Failed replacement
still preserves the old copy.
Interrupted ranged transfers retain their partial bytes for the next call; rejected or unsatisfiable ranges still restart safely.
A transient identity-probe failure preserves completed downloads for retry. Date-bound resumes can adopt a strong response tag without re-downloading.
Fresh HTTP responses no longer inherit stale HEAD validators. Changes in ETag strength trigger updates even when Last-Modified is unchanged.
Incomplete downloads report the received byte count and clear temporary bytes and completion metadata before a retry.
dp_connect(config = ...) and
options(datapond.duckdb_config = ...) allow memory, thread
and spill limits before database attachment.
Weak ETags no longer enable a ranged resume. Update and partial-download identities are scoped to the source URL, so a new mirror cannot accidentally match an unrelated ETag.
Completed downloads retain their metadata until installation succeeds. If installation fails, a retry reuses the completed file after rechecking its URL, revision and size, instead of downloading it again.
A failed database validation clears the completed-transfer marker so retry fetches fresh bytes. A full fresh response uses its own identity without a redundant transfer; resumed responses still restart after validator changes.
Empty quoted ETags remain valid strong validators, and differing weak ETags trigger an update even when Last-Modified is unchanged.
DuckDB configuration accepts scalar numeric and logical values as well as strings, with clear errors for missing or vector values.
Installation documentation now links the live CRAN release (0.1.0, published September 27, 2026) and distinguishes development-only changes.
The identity probe used when a server’s HEAD carries no validator
reads one byte through a connection and closes it, so a server that
ignores Range can no longer make the client buffer the
whole database in memory just to read its headers. Covered by loopback
tests where HEAD answers 405 and where Range is ignored
(with a memory-profiling assertion).
Real downloads work again: the response headers curl returns are
a vector of lines, not one string, and a resumed transfer’s
Content-Length is only the remaining range (the total comes
from Content-Range). A server that answers a stale
If-Range with the whole file makes curl abort the resume;
the client now discards the partial file and restarts. These paths are
covered by a real loopback HTTP server test (httpuv + callr,
Suggests).
dp_download() streams to
<file>.part, resumes only while the remote file is
the same revision (a strong validator, ETag else Last-Modified, recorded
beside the partial file; the ranged request is bound to it with
If-Range and the GET response’s validator is checked, so a
file that changes between HEAD and GET is re-downloaded from scratch; no
validator, no resume), validates the finished file as a DuckDB database,
rejects a directory at the destination, and errors instead of warning
when the destination cannot be replaced. An existing local copy is never
damaged by a failed download.
dp_update() compares the remote file’s ETag (else
Last-Modified) with the identity recorded at download time
(<file>.datapond.json) instead of the file’s
modification time; equal sizes alone never count as current; the
registry is re-fetched first.
The registry response is decoded as UTF-8 explicitly, so the package works in a C/ASCII locale; an unparseable response falls back to the cached copy.
dp_download(path = "some/dir/") creates the
directory.
dp_list(),
dp_databases(), dp_info(),
dp_registry(), dp_connect(),
dp_disconnect(), dp_download(),
dp_update(), dp_describe(). Mirrors the Python
datapond package (0.1.3) against the same registry and
Hugging Face-hosted DuckDB files.