Hesela / Research library
Research & analyses
Storage engineering, examined through sources, measurements and reproducible methods.
154 analyses · Page 1 of 6
Newest firstCache stampedes: early refresh still needs a concurrency policy
A worked XFetch example separates the chance of an early refresh from duplicate-work suppression, with explicit limits for local coalescing and cache refill tokens.
Time-series rollups: equal summaries, different data
BTrDB separates versioned changes from aggregate queries. Exact examples show why matching rollups can hide changed samples and misweighted averages.
Hyperbolic caching: priorities can cross without hits
An exact two-object example shows why frozen cache priorities can choose the wrong candidate, and why eight random draws do not guarantee the global minimum.
Recovering bzip2 blocks: which bytes survive corruption?
Controlled bzip2 experiments show which fragments survive damage and why intact blocks can change order while every stored CRC still passes.
Zarr archive checks: when missing chunks become plausible numbers
Controlled Zarr and Xarray experiments trace deleted chunks, valid old shards and stale metadata through successful reads and misleading aggregate values.
Interlaced recording: two neighbors, different write costs
Compare IMR track overlap, memory versus disk-backed preservation, and byte write amplification with six reproducible accounting examples.
PostgreSQL TOAST: which checks reach the value?
A controlled one-bit change survives row counts, length and prefix reads; compare page checksums, structural checks and trusted payload digests.
Full-path indexing: why renaming can move the index
How namespace layout changes rename cost, how BetrFS range rename addresses it, and what a fast rename leaves unproven.
LTFS performance: when does the write job finish?
Understand LTFS write completion, index recovery and archive verification: partition-switch costs, salvaged records and the limits of stored file hashes.
Magnetic grain models: what the tiles actually represent
A source review separates model tiles from fabricated media and explains why grain proportions, sampling and decoder stopping rules matter.
Reading an SD card label: capacity, speed and the host
Separate SD capacity families, bus ceilings and performance classes, including the A2 command-queue exception and the version-specific 1.8 W power limit.
Restoring file names: when matching hashes hide a lost entry
Unicode normalization, case folding and lossy decoding can merge file names. A controlled experiment shows why matching payload hashes cannot verify a restored namespace.
Rsync after a one-byte shift: less payload, not free work
Native rsync experiments distinguish shifted-block reuse, file selection and interrupted in-place retries that exit successfully without restoring matching bytes.
Sharing an SSD: who pays for the slowdown?
Assess SSD slowdown fairness, host-scheduler limits and NAND suspension, distinguishing hardware observations from simulated policies.
SQLite imports: where exact identifiers lose digits
Controlled SQLite and GNU sort experiments separate rounded identifiers, discarded records and lost telemetry status from successful integrity checks.
CRC-32 vs CRC-32C: which bit errors escape detection?
Reproducible bit-flip examples distinguish CRC-32 from CRC-32C, burst span from error count, and message-length guarantees from collision probabilities.
Restoring Linux ACLs: identical bytes, different permissions
A controlled Linux probe traces how rsync options, chmod, mask recalculation and default ACLs change restored permissions while payload hashes remain identical.
What inotify misses: mapped writes and collapsed change events
Measured Linux boundaries: mapped writes, shared watch ownership, rebound fsnotify event names and directory scans that miss live files.
DNA storage: coverage, dropout and verified recovery
Distinguish missing strands from deleted bases, then trace DNA coverage through independent equations, decoder stalls and verified recovery.
Encrypted deduplication: who can test a file guess?
Compare encrypted deduplication, client ownership proofs and remote storage audits, including why a 460-block sample can miss a single lost block.
FITS archive checks: reordered pixels can preserve both checksums
FITS checksums can pass after pixels or saved values change. Reproducible Astropy and CFITSIO evidence shows why exact values, null masks and inventory need separate checks.
FUSE writeback caching: reads behind writes and the fsync contract
How FUSE caching changes requests, why a write-only open can require reads, and how Linux handles an unimplemented fsync. A version-pinned source review.
HDF5 dependencies: missing sources and truncated raw data
HDF5 experiments show missing VDS sources returning fill values and truncated external raw files changing integers despite a configured fill sentinel.
Keeping Btrfs incremental backups recoverable
Preserve parent and clone-source identity, separate stream checks from restoration, and protect the receive-time write window in Btrfs replication.
LZ4 frame checks: block order, dictionaries and completion
Reproducible LZ4 experiments separate block integrity, dictionary selection, restored-byte identity and frame completion.
Ordering storage generations across counter wrap
A reproducible counterexample connects finite sequence numbers, JBD2 journal comparisons and PostgreSQL XIDs: local ordering needs a bounded history.
QCOW2 backups: an unchanged overlay can expose different data
QCOW2 backing and external data files can change guest bytes while metadata hashes and consistency checks pass. Native experiments separate these guarantees.
What a recovered disk image can actually prove
Twelve ddrescue cases and six GPT fixtures show what recovered images prove; native parser warnings survive a successful exit status.
Where Linux charges I/O: cache reads, mapped writes and cancelled writeback
Measured Linux I/O accounting, plus TMO source evidence: why fewer swap-ins can accompany worse throughput and how to compare PSI observation windows.
Why SQLite index results can change when database bytes do not
SQLite experiments trace stale indexes and cascading deletes that intact backups and passing checks do not reveal.
Looking for a technology? Browse technology cards or open the knowledge corpus.