OpenDAL: open files from remote storage (S3, GCS, Azure, HTTP)

[anitnilay20]

Problem

Thoth currently opens files from the local filesystem only. Data teams routinely work with files stored in object storage (S3, GCS, Azure Blob) or accessible via HTTPS. Opening those files today means downloading them manually first.

Solution

Integrate OpenDAL — a unified storage abstraction layer that supports 40+ storage backends behind a single Operator API. One dependency, many backends.

Supported backends (initial scope)

Backend	URI scheme	Notes
Local FS	file://	Replaces current path handling
S3 / R2	s3://bucket/key	AWS, Cloudflare R2, MinIO
GCS	gs://bucket/key	Google Cloud Storage
Azure Blob	az://container/blob
HTTPS	https://host/path	Public URLs

User flow

File → Open Remote… (or paste URI into open dialog) → credential prompt if needed → file streams into Thoth exactly like a local file.

DuckDB integration (#148) works transparently: for formats DuckDB reads natively, the host downloads to a temp path and hands it to DuckDB. For plugin-loaded formats, OpenDAL streams bytes through the existing get-range path.

Credential management

Credentials stored per-backend in the OS keychain (same secure-storage infrastructure already in place). Users configure once; Thoth remembers.

Changes

• Add opendal crate to Cargo.toml with relevant feature flags
• src/storage/ — StorageBackend abstraction wrapping OpenDAL Operator
• File open dialog: accept URIs in addition to local paths
• Credential UI: simple key/value form per backend type (access key + secret, service account JSON, etc.)
• Progress indicator for remote downloads (size may be unknown)

Acceptance criteria

• Open files from S3, GCS, HTTPS URIs via the open dialog
• Credentials stored in OS keychain, not plaintext config
• Progress bar during download (indeterminate if size unknown)
• Local file path handling unchanged (no regression)
• DuckDB native-format path works: remote file downloaded to temp, DuckDB scans it
• Large files: stream into temp file rather than full memory load