Agave v4.4.0-rc.1 Carries the Fix Marinade Ties to Skipped Alpenglow Leader Slots

Anza tagged Agave v4.4.0-rc.1 on 26 October 2026, a release candidate for Anza's Solana SOL$101.56+0.8% validator client that includes a backported fix to how block producers release old chain state between turns. Marinade MNDE$0.025+0.9% says that fix cures a bug in Agave v4.3.2 that has had validators skipping about half the slots in their first leader window of every epoch since Alpenglow went live on mainnet. Until a release carrying the fix ships, Marinade has stopped charging validator bonds for those slots.

The v4.4.0-rc.1 release page calls it a "Mainnet-beta Upgrade Candidate release recommended for 10% adoption", and also recommends it for testnet and devnet. That makes it a staged candidate for a slice of mainnet stake; Anza has not recommended it to every operator. Agave v4.3.2, which Anza named the recommended release five days before the Alpenglow switch, is still the latest stable version on Anza's releases page.

Marinade's skipped-slot figures for Agave v4.3.2 under Alpenglow

A leader window is the short run of consecutive slots in which the network schedules one validator to produce blocks. A skipped slot is one where the scheduled leader produces no block, so it collects nothing for that slot, and stake programs that score uptime count it against the validator.

Marinade's numbers come from its own analysis of the chain. In a validator-bonds pull request merged on 26 October, Marinade wrote that in epoch 1068 nodes running Agave 4.3.2 skipped "51.9% of first-window slots, against 0.05% of other slots", while Firedancer nodes skipped none. Epoch 1068 began at 19:28 UTC on 25 October, according to Solana Compass epoch data. The 51.9% figure is Marinade's alone; no other party has published a matching measurement, and Anza has made no public statement about the skips.

Marinade put the finding in plainer terms in an X post at 10:01 UTC on 27 October: "Agave 4.3.2 nodes on Alpenglow skip about half of their first leader window slots each epoch. The fix is merged but not released yet." The post went out the morning after the release candidate was tagged and does not mention it, which fits rc.1's status as a pre-release.

How the Agave PoH fix in v4.4.0-rc.1 works

The change is Agave pull request #16302, "poh: cache start bank fields so pruned banks are freed", merged on 22 October. Proof of history (PoH) is the running clock a Solana leader uses while it builds blocks, and each reset of that clock starts from a "bank", Agave's in-memory record of chain state at a given slot.

According to the PR, Agave's PoH recorder kept a reference to its starting bank, so the bank could not be freed after the validator had otherwise discarded it. Under Alpenglow, the PR says, PoH resets only when a validator's leader window starts. Between windows the old bank stayed in memory, holding about 1 GB of stake data on mainnet, and it was finally dropped at the next window start, on the block-production thread, while the PoH lock was held. The fix keeps only four small fields from the bank (its slot, bank ID, nanoseconds per slot and leader ID), so Agave can free the rest on schedule.

In local test-cluster runs posted in the PR, the median PoH reset for banks 400 to 800 slots old fell from 30.5 milliseconds to 27 microseconds. Those figures come from a test cluster; the PR reports no mainnet measurements.

The change was backported to the v4.4 branch as #16312 on 23 October. Approving it, AshwinSekar called it a "mechanical change" that lets the bank "be async dropped during root cleanup, rather than holding onto it till the next leader window and performing a sync drop during leader's build time."

Neither Agave PR mentions skipped slots. The link between this memory fix and the first-window skips comes from Marinade, whose ds-sam pull request of 23 October says Agave 4.3 to 4.5 skip "most of each validator's first leader window every Alpenglow epoch" and cites agave#16302 as the still-unreleased fix. The reviewer's description of a synchronous drop during the leader's build time is consistent with Marinade's account, but Anza has not publicly confirmed the cause.

Marinade's validator bond waiver for epochs 1065 to 1070

Marinade runs two systems the skips feed into. Under Protected Staking Rewards (PSR), validators post a SOL bond, and when a validator earns its stakers less than expected, Marinade pays the shortfall to stakers from that bond in a settlement at the start of the next epoch, with a 1% grace on downtime, per its bonds and settlements documentation. The Stake Auction Marketplace (SAM), where validators bid for Marinade's delegated stake, requires an uptime score of at least 0.8, per the ds-sam PR.

That ds-sam change, merged on 23 October, removes up to four first-window leader slots per validator from the uptime and PSR calculations for epochs 1065 to 1067. Without it, Marinade wrote, small validators could drop below the 0.8 eligibility threshold. The 26 October validator-bonds PR extended the window to epochs 1065 to 1070 and synced the same list across Marinade's validator-bonds, institutional-staking and ds-sam code. Epoch 1070 had not started when that change merged; Solana Compass epoch data puts its start at about 22:30 UTC on 27 October. Marinade has not said what it will do if operators are still on v4.3.2 after epoch 1070.

Marinade is both the party measuring the skips and the operator of the penalty system it is waiving. Its X post gave the reasoning in one line: "Bonds are for a validator's own mistakes. A client bug isn't one." The trade-off is that Marinade's stakers forgo the bond settlements those skipped slots would otherwise have triggered, while small Agave operators stay eligible for Marinade stake as the fix moves through release.

What Agave v4.4.0-rc.1 means for Solana validator operators

For now, the fix reaches only operators who opt into a release candidate aimed at 10% of mainnet. The release carries other backports too, including dedicated jemalloc memory arenas for replay and a streamer change that allows 3,000 unstaked connections with unstaked traffic throttled to 100 TPS. Anza CEO Brennan Watt posted about the release on 26 October: "publishing Agave v4.4 release before a new RUSTSEC pops challenge (impossible)".

No v4.3 backport appears in either Agave PR, and Anza has not said whether one is planned. Until a stable release carries the fix, Marinade's waiver is the only published adjustment for operators losing first-window slots. Marinade's count of zero first-window skips on Firedancer nodes, against roughly half on Agave 4.3.2, is a client-diversity data point from Alpenglow's first week on mainnet.