direct-recovery · research · also Distributed error Isolation and RECovery Techniques

DIRECT recovery

DIRECT is a research approach that repairs small corrupted regions from distributed replicas while preserving the storage system's update order.

Related Observatory article: Read on hesela.com

DIRECT combines fine-grained data recovery, local metadata redundancy and recovery ordered with writes. Logical replicas can have different physical file layouts, so a damaged block cannot simply be replaced by the same block from another replica. [1]

Its key-value implementation places a patch request in its own Paxos log entry. The donor supplies the preceding log-prefix state; installation is atomic. Keys absent from that complete range are implicit deletions, including deletions whose tombstones have already been compacted away. Affected historical snapshots need separate handling. [1]

An empty complete range and an incomplete scan are different inputs. Only the former supports removing every local key in that range. The executable example compares range replacement with a merge that preserves a deleted key. [3]

Evaluation used injected errors, not a field trial of extended SSD lifetime. [1] An earlier eight-system study independently showed that replication alone does not guarantee recovery from a single filesystem fault. [2]

Relations

related
Read repair, Silent data corruption, State-machine replication, LSM tombstone, Sorted string table

Limitations

  • Recovery depends on identifying the affected logical range and obtaining an intact donor at the required consistency point. [1]

Citations

  1. Tai et al. Who's Afraid of Uncorrectable Bit Errors? Online Recovery of Flash Errors with Distributed Redundancy. USENIX ATC 2019, sections 3-5. (accessed 2026-11-22)
  2. Ganesan et al. Redundancy Does Not Imply Fault Tolerance. USENIX FAST 2017. (accessed 2026-11-22)
  3. Hesela: complete-range patch counterexample, executable model and tests (accessed 2026-11-22)
Sources & caveats
  • Primary source review, not a DIRECT implementation, hardware experiment or benchmark replication. [1]
  • The linked example is an in-memory teaching model. Its complete-range and log-position assertions are assumptions supplied by the caller, not a distributed proof. [3]

JSON: /data/terms/direct-recovery.json

Published by Hesela team · Updated

Report a correction · Editorial standards · Reuse guidance

Cite this page
Hesela team. DIRECT recovery. Hesela. 2026-11-22. https://hesela.dev/terms/direct-recovery/