---
title: "Erasure Coding: Why Two Lost Chunks Are Not Always Two Lost Hosts"
description: "Leviathan Systems — Technical content Published 2026-10-21; updated 2026-10-21."
canonical: "https://www.leviathansystems.co/articles/erasure-coding-chunks-host-failure-placement"
last-updated: "2026-10-21"
---

# Erasure Coding: Why Two Lost Chunks Are Not Always Two Lost Hosts

Leviathan Systems — Technical content
Published 2026-10-21; updated 2026-10-21.

Translate each requested host-loss scenario into unavailable chunks. In the fictional 3+2 example, any three of five healthy, matching chunks can reconstruct the object. Compact placement A puts those chunks on three hosts; every two-host loss leaves fewer than three. Spread placement B uses five hosts and retains three after any two host losses. B meets that stated coding-survival condition under the example's assumptions. Neither result establishes active I/O or application recovery.

## Define what the 3+2 claim actually promises

Ceph Squid's erasure-code documentation describes K data chunks and M coding chunks. Its named 3+2 example relates two tolerated losses to OSDs, the object-storage daemons holding the chunks. The erasure-code profile and Jerasure references separately describe chunk counts, coding techniques and placement failure domains.

The buyer's question here is narrower than a complete storage-availability assessment: does the offered placement retain enough chunks after any two hosts become unavailable? Answering it requires both the applicable code's reconstruction property and the mapping from chunks to hosts. The numbers 3+2 alone do not supply that mapping.

Fictional case EC-25 stipulates one immutable object generation encoded into D0, D1, D2, C0 and C1. These are five distinct, valid chunks; any three are sufficient to reconstruct the object in this Reed-Solomon-style case. Each chunk occupies a different OSD. Start with all five healthy and available, with no other losses, corruption or mismatched generations. These premises define the example rather than a guarantee for every coding technique.

The three data chunks collectively represent the original data. The two coding chunks provide additional reconstruction information. They are not five complete copies, and the required surviving three need not be the three original data chunks.

References: [C1](https://docs.ceph.com/en/squid/rados/operations/erasure-code/), [C2](https://docs.ceph.com/en/squid/rados/operations/erasure-code-profile/), [C3](https://docs.ceph.com/en/squid/rados/operations/erasure-code-jerasure/)

## Hold the code fixed and change only host placement

Offer A supplies this compact mapping: host H1 holds D0 and D1; H2 holds D2 and C0; H3 holds C1. Although the five OSD identities differ, H1 and H2 each contain two of the object's chunks.

Offer B uses the same five chunks and reconstruction premise. It places D0 on H1, D1 on H2, D2 on H3, C0 on H4 and C1 on H5. Each host holds one chunk of this object.

The offers therefore have identical coding arithmetic but different correlated-loss exposure. When a host becomes unavailable, count every chunk on that host. A separate OSD identity cannot keep a colocated chunk available through the host's outage.

A is an abstract supplied OSD placement with host colocation. It does not satisfy a Ceph placement rule that prevents two chunks of this object from sharing a host.

References: [C3](https://docs.ceph.com/en/squid/rados/operations/erasure-code-jerasure/)

## Evaluate all three compact host pairs

In A, losing H1 and H2 removes D0, D1, D2 and C0. Only C1 remains: one chunk, below the required three.

Losing H1 and H3 removes D0, D1 and C1. The survivors are D2 and C0: two chunks. Losing H2 and H3 instead leaves D0 and D1: again two. Those are all three possible two-host combinations in A, and each leaves an insufficient surviving set.

One lost host can have a different result. Losing H1 alone leaves D2, C0 and C1. Those three satisfy the stipulated reconstruction condition even though two are coding chunks. A can therefore meet this particular single-host coding condition while failing the requested two-host condition.

In B, any two different hosts remove exactly two chunks. Three remain, whichever pair is unavailable, so B meets the requested coding-survival condition from the stipulated healthy starting state. This is a deduction from the fictional placement, not a failure-test result or a claim that reconstruction has run successfully.

The failed condition in A establishes insufficient listed survivors. It does not establish permanent destruction of unavailable chunks or rule out a separately retained recovery copy.

## Include chunks already unavailable before the next loss

Now change the starting condition: C1 is already unavailable. In A, an additional H1 outage removes D0 and D1, leaving only D2 and C0. The earlier conclusion that H1 alone leaves three relied on C1 being available.

Count distinct unavailable chunks across the whole scenario. A chunk already unavailable and also located on a newly lost host counts once, not twice. In EC-25, five minus that distinct unavailable count must be at least three for the stipulated reconstruction condition to hold.

The same starting-state issue applies to B. If C1 is already unavailable and H1 and H2 then become unavailable, only D2 and C0 remain. B's healthy-start two-host result does not establish survival of those two additional host losses during that degraded condition.

Have the supplier state whether the requirement covers two host losses from a healthy state or additional losses while data is already degraded. Those are different requested conditions.

## Connect the failure-domain rule to the offered mapping

Ceph Squid's Jerasure documentation describes a selected CRUSH failure domain such as host or rack: the placement rule prevents two chunks of the same object from sharing that selected bucket. A host-level rule therefore addresses a different boundary from merely selecting different OSDs.

Five chunks with no two on the same host require at least five distinct eligible hosts for that placement. Three hosts cannot meet that condition by containing more OSDs. Conversely, a count of five hosts somewhere in the cluster does not establish that this object's five chunks are spread across them.

Request the applicable coding profile, coding technique, pool's actual placement rule and relevant chunk-to-OSD-to-host mappings. Ceph Squid's pool documentation identifies CRUSH rules as governing chunk placement. Read the offered pool's rule together with its topology and mapping; the profile's coding counts alone cannot prove host separation.

A dataset contains more than the single object illustrated here. The supplier's support must cover the objects or placement sets needed for the offered dataset claim; a favorable mapping for one selected object cannot establish the condition for all of them.

If the requested loss boundary is a rack, use the rack membership instead. Different hosts can share a rack or another common dependency. Neither the host count nor the coding label establishes independence from shared power or network failures. Those conditions require the actual dependency mapping and an explicit requirement.

References: [C1](https://docs.ceph.com/en/squid/rados/operations/erasure-code/), [C3](https://docs.ceph.com/en/squid/rados/operations/erasure-code-jerasure/), [C4](https://docs.ceph.com/en/squid/rados/operations/pools/)

## Three reconstructable chunks can still fall below an I/O prerequisite

Retaining three valid chunks answers the stipulated coding question. A running storage service has separate conditions for accepting I/O.

Ceph Squid's pool documentation describes min_size for an erasure-coded pool as the minimum number of active shards needed for a placement group to become active. That service prerequisite is distinct from K, the data-chunk count. Meeting a reconstruction count alone cannot establish that the placement group is active or that clients can use it.

For a separate illustration, stipulate a service that requires four active shards. After two host losses in B, three valid chunks remain and are stipulated active. Three is sufficient for EC-25's algebraic reconstruction but less than this service's four-shard prerequisite. The coding condition is met; the independently stipulated active-shard requirement is not.

Four is an invented comparison input, not a recommended setting. The example determines no real pool's operating outcome. Evidence about current cluster state, service configuration and client access must support an actual availability claim; recovery time and application return need their own evidence.

References: [C4](https://docs.ceph.com/en/squid/rados/operations/pools/)

## Ask for a placement answer and a separate service answer

For this buyer's stated healthy-start requirement, A does not preserve enough chunks after any two host losses. B does, under EC-25's premises. That is the resolved comparison; neither conclusion selects a complete storage architecture.

A useful supplier return connects the named code and version to the offered placement, then identifies the minimum surviving valid chunk count across the requested host-loss combinations. It states the starting condition and any additional unavailable chunks. For the dataset claim, it explains how its evidence covers the relevant placements rather than presenting one favorable object.

Read any broader availability promise separately. What keeps the storage service active under the same condition? What client access is covered? What restoration or recovery result is promised, on what evidence? A statement that the object is mathematically reconstructable cannot supply those answers.

The raw, protected and client-usable capacity guide addresses the separate capacity purchase. The checkpoint-versus-snapshot guide addresses which application state is retained. Enough surviving chunks cannot establish either the buyer's allocation size or the desired training generation.

## Carry an approved physical layout into the installation handoff

This comparison alone creates no requirement to buy more hosts or change an installation. If the selected storage plan includes physical mounting or cabling work, its approved host identities, rack positions and connections become useful installation inputs.

Leviathan's rack-and-stack service covers agreed equipment mounting, IT cabling, labeling and physical handoff records. The related service and project inquiry provide a place to discuss that work. Share a nonconfidential equipment summary, released rack and connection plans, site constraints and required records.

The appointed storage owner retains responsibility for coding configuration, logical placement and storage-service decisions. Physical installation records can support the identity and connection map within their agreed scope; they do not establish the coding-survival or application-recovery result.

References: [L1](https://www.leviathansystems.co/services/rack-and-stack)

## Primary sources

- [C1 — Ceph Squid: Erasure code](https://docs.ceph.com/en/squid/rados/operations/erasure-code/). Checked 2026-10-21. Versioned concepts and named 3+2 OSD-loss example, with rack separation established by a CRUSH rule. Supports separating the coding allowance from the chosen physical failure domain.
- [C2 — Ceph Squid: Erasure code profiles](https://docs.ceph.com/en/squid/rados/operations/erasure-code-profile/). Checked 2026-10-21. Defines K data chunks and M coding chunks of equal size. Useful for identifying the coding scheme before evaluating a host-loss requirement.
- [C3 — Ceph Squid: Jerasure erasure code plugin](https://docs.ceph.com/en/squid/rados/operations/erasure-code-jerasure/). Checked 2026-10-21. Documents coding techniques and CRUSH failure domains, including prevention of two chunks sharing the selected host or rack bucket. Technique-specific documentation must support the actual offered code.
- [C4 — Ceph Squid: Pools - min_size](https://docs.ceph.com/en/squid/rados/operations/pools/). Checked 2026-10-21. Describes the erasure-coded pool's active-shard prerequisite for placement-group activity. Supports the distinction from a reconstruction count; supplies no setting recommendation or observed availability result here.
- [L1 — Leviathan Systems: Rack-and-Stack and Colocation Installation](https://www.leviathansystems.co/services/rack-and-stack). Checked 2026-10-21. Existing physical mounting, IT cabling, labeling and handoff scope, with project inquiry. Storage configuration and application recovery have separate appointed owners.

## Related Leviathan resources

- [Raw, protected and client-usable storage capacity](/articles/gpu-storage-raw-versus-protected-usable-capacity)
- [Storage resilver and scrub: distinguish restoration from verification](/articles/openzfs-resilver-versus-normal-checksum-scrub-scope)
- [Training checkpoints and storage snapshots: identify the retained state](/articles/gpu-training-checkpoint-versus-storage-snapshot)
- [Rack-and-stack physical installation scope](/services/rack-and-stack)
- [Discuss an approved physical storage installation package](/contact?service=rack-and-stack&from=%2Farticles%2Ferasure-coding-chunks-host-failure-placement)

## Need physical installation records for the selected storage layout?

Share the equipment identities, approved rack and connection plans, site constraints and required physical handoff so we can discuss scope and availability.

[Discuss the scope](/contact?service=rack-and-stack&from=/articles/erasure-coding-chunks-host-failure-placement)
