EIP-8188

State Tiering by Write Age

Proposed for Inclusion in Hegotá

Wei Han Ng (@ngweihan_eth)

Ethereum Foundation (Stateless Consensus)

Not All State is Equal

Majority of the state is inactive

One protocol-level signal: `last_written_period`

1 · Period counter

period = (block - PERIOD_START) // PERIOD_LENGTH

~6 months per period. Derived from block number, no extra state.

2 · Add period metadata for accounts and storage slots

Account RLP gains a 5th field. Storage slots wrap their value in a list with the same field. Legacy entries decode as last_written_period = 0.

3 · Write cost changes

Writing to Inactive state costs more. Writing to Active state costs less.

What bumps the period

Reads are unchanged.
Only actual state mutations bump the period.

Architecture unlocks for EL clients

LSM clients Geth · Nethermind · Besu · Ethrex

Extract Inactive subtrees to an append-only flat file, leaving small stubs in the live LSM. Compaction sees fewer files, working set shrinks.

Erigon MDBX + segment domains

Fits very well with this EIP. Already has hot-cold storage separation where items are moved into immutable flat files periodically.

Reth MDBX + Nippy Jar

Split plain/hashed/trie state along the Active/Inactive axis.

Separate cold subtrees out of the live LSM

Geth prototype · 1 year of mainnet replay

Trie Nodes

148GB→33GB

−78%

Live LSM after extracting Inactive trie nodes

Replay scope

2.6M blocks

head 19,999,256 → 22,627,956 (~1 year)

Stub durability

80% survived

Original stubs unchanged across the entire import

Performance benchmarks: TBD · Prototype: github.com/weiihann/go-ethereum · branch feat/eip-8188/inject

FAQ

Doesn't the metadata bloat the state?

Worst case ~3.4GB (0.36GB accounts + 3GB slots). Pre-fork untouched state pays zero. Geth prototype measured only +300MB.

Why not state expiry?

State expiry needs decentralized resurrection infrastructure. None of which exists yet. EIP-8188 has no such prerequisite — state stays accessible, writes just cost more.

Is it compatible with future tree migration?

Compatible. When the trie moves to a page-based layout (e.g. EIP-7864), the period aggregates to the page: page.period = max(slot.period).

Why Hegotá, not later

Small, contained EIP. Two RLP encoding changes. No tree migration.
Metadata accumulates as early as possible while clients work on database optimizations.
If pricing complexity is a concern, defer pricing to a later fork.
One possible alternative: record metadata as block number in Hegotá and decide the gas-pricing mechanism later. Costs ~4× more bytes than periods, but allows us to be more flexible.