DataBucket
Page framing, row directories and index storage for WorkTable.
Version 3 cutover
DataBucket 0.7 writes page format 3. Older page formats are rejected with a version error; opening them does not convert or remove them. Deployments that can regenerate their data should explicitly recreate the store. Retained data requires an application-specific conversion with the old reader.
For a page of P bytes, the general header occupies the first 28 bytes. Data
pages contain row bytes followed by a live-row directory at the page tail.
Each directory entry is a little-endian (u32 offset, u32 length) pair; its
offset is relative to the payload. A CRC-32 is stored at P-8 and the entry
count at P-4. The checksum covers the entire payload except its own word,
including padding, directory and count. Header identity and version are
validated separately. Index and metadata pages retain their existing body
layouts, with the new page-version marker.
data_page_row_capacity computes a row allocation budget that reserves the
maximum directory space for the minimum archived row size. DataPage::encode
and decode validate directory extents and integrity. Updates, deletion and
relocation must maintain DataPage::rows; raw row bytes alone are not a
complete v3 persisted data page. The file-level update_at accepts a readable
and writable file, validates the existing live row and updates the checksum.
The table schema version and crate package version are separate from the page-format version. WorkTable's Vec snapshot files use a different container and cannot be treated as ordinary DataBucket space files.
CRC validation detects damaged data pages; it does not provide a transaction log, atomic multi-file commits or crash repair. WorkTable owns synchronization and durability policy.
Storage domains
storage::StorageDomain groups every table and index page in one database-wide
generation. begin_generation() returns a builder for page puts, page deletes
and semantic catalog changes. commit_generation() stages immutable page
objects, prepares the system catalog, writes its checkpoint, conditionally
advances the small bootstrap head, and only then publishes the prepared catalog
in memory. A failed stage or conflicting head leaves the preceding catalog
visible.
DataBucket defines the bounded catalog records and the SystemCatalog provider
interface. WorkTable supplies the provider as a generated table. This keeps the
Cargo dependency one-way while DataBucket retains the only catalog write permit
at runtime. Applications query WorkTable's read-only catalog view.
storage::PageStore supplies restart, range-read, stage and conditional commit
operations. With s3-support, storage::s3::S3PageStore implements that
contract for Tigris and compatible S3 services. Data pages are stored in
content-addressed segments. Adjacent dirty pages may coalesce up to 4 MiB, but
the target is never a minimum. One changed page remains one page-sized data
upload.
The generated catalog checkpoint is also divided into content-addressed
PAGE_SIZE chunks. A mutation uploads only changed catalog chunks, a compact
chunk manifest and the bootstrap head. Catalog rows use stable tombstones so a
delete does not shift every later row and turn a small mutation into a complete
catalog upload. StorageDomain::open restores the committed catalog before any
page lookup, and read_page resolves a logical page through that catalog and
issues one bounded range read.
The library is no_std with alloc, including its default validate-reads
feature. It uses Nagoya's portable I/O traits and requires an allocator. The
concurrent index dependency uses OS services through libc on supported targets;
no_std does not imply a bare-metal implementation. The command line tools and
host tests use std.
Run sh scripts/check-no-std.sh -p data_bucket --lib to verify the default
library graph with Rust std removed from the target sysroot. Build scripts and
proc macros retain their normal host environment.
Command line tools
Create a sample file containing 2,500 records. An existing file is refused:
Inspect page identities and live row extents without consulting an index:
--count is a record count. The dumper reports each row's payload offset,
archive length and absolute file offset. --hex includes its archive bytes;
typed deserialization requires the owning application's row schema. For a
nondefault stride, supply --page-size; supported values are 512, 4096, 8192,
16384 and 32768 bytes. The tools validate data-page checksums and directories.