Yunagi/Notices
案YG-NTC-04 · 2026.08
案 Notices
Twenty-one entries, February 2026 to the two windows we have not worked yet. Price changes, new capacity, maintenance, and the incidents — written the day they happened, kept whether or not they flatter us.
Twenty-one entries run from Seoul opening in February to the two maintenance windows we have not yet worked. Six tags: Promotion, Product, Network, Maintenance, Incident, Security. Choosing one rewrites the address bar, so a filtered view can be sent to a colleague, and a link to a single entry opens with that entry's tag already selected. Every price shown struck through elsewhere on this site points back at the entry here that changed it. Nothing is edited after posting; corrections are appended and carry their own date.
記録 The Archive
Entries are grouped by month, and the month heading stays at the top of the screen while you read beneath it, so the date is never more than a glance away. Each entry has a permanent anchor built from its date and a slug — 2026-07-11-hkg-transit-withdrawal — and that is the form our support replies quote. Scheduled work sits above today's line with a vermilion punch mark against the date, the same mark the cut leaves. When the window closes the mark goes, the entry stays, and the outcome is written into the body.
2026.09.13
#
Maintenance
Sunday 13 September, 02:00–06:00 UTC. One of four 100G ring cards at LHR-1 is replaced. Traffic drains onto the HKG–LHR span before work starts and the ring runs on three cards for the window. Expect no loss and up to 9 ms of extra round-trip on London routes. No instance reboots. The all-clear is posted here.
LHR-1
2026.09.06
#
Maintenance
Sunday 6 September, 03:00–09:00 UTC. Hosts take a kernel carrying a fix for an NVMe queue stall. VPS and VDS are live-migrated: expect one pause of under two seconds. Dedicated chassis are untouched — you reboot those yourself, any time in the following fortnight. Per-instance windows appear in the console seven days ahead.
JFK-1 · LAX-1
2026.08.14
#
Promotion
Every VPS tier at HND-1 and SEL-1 falls to launch pricing, from $4.20 a month against $6.00. The rate is attached to the instance, not the first invoice: it holds until you destroy it, and survives resizes within the VPS range. The other four metros stay at $5.20–$5.90. No coupon, no minimum term, no card needed to see it.
HND-1 · SEL-1
2026.08.07
#
Product
SEL-1 opened in February with VPS only. VDS and dedicated chassis are now in stock there at the same $18–$22 and $85–$99 the other metros charge, on the same five working day lead time. Seoul has no GPU row; those stand only at Tokyo and Hong Kong. HND–SEL round-trip is 34 ms, SEL–HKG 62 ms.
SEL-1
2026.08.01
#
Promotion
Registration or transfer of a .cloud name is $9.50 for the first year and $11.90 at renewal. Both numbers are printed beside each other at checkout, because the second is the one you pay five times. No volume requirement and no rebate paperwork. .com is unchanged at $10.20, renewing at $10.20.
All six
2026.07.28
#
Security
The console accepts WebAuthn security keys as a second factor, or as the only factor where you prefer that. API tokens can be scoped to one product and one metro and expire on a date you set. Existing unscoped tokens keep working until 2026.11.30 and must be re-issued after it. All sessions are cut on password change.
All six
2026.07.20
#
Product
Our own image model leaves preview. Every account includes 1,000 renders each month; beyond that it is $0.012 per image, metered per render rather than per second. The endpoint keeps its OpenAI-compatible shape, so existing tooling works unchanged, and preview keys keep working. Prompts and outputs are not retained after the response is returned.
HND-1 · HKG-1
2026.07.11
#
Incident
An automated rebuild of our prefix lists dropped the HKG-1 /22 from what we announce to three of four transit partners. From 14:06 to 14:37 UTC roughly 70% of inbound paths to Hong Kong went dark. Instances kept running and the private ring was unaffected, so traffic to the other five metros never stopped; nothing rebooted and nothing was lost. Cause: the rebuild read a stale inventory export, and we pushed it to all four edges at once. Since then a prefix list that shrinks by more than one prefix is refused without a second reviewer, edges take changes one at a time ten minutes apart, and an off-network probe alarms when any metro loses a transit path. Credits were applied on 12 July.
HKG-1
2026.06.30
#
Network
HKG-1 now carries the VDS tier at $18–$22, alongside VPS, dedicated, and the GPU row. Dedicated chassis lead time falls to five working days at all six metros, from ten: we hold spare chassis in each cage rather than shipping them from London. HKG–HND is 51 ms, HKG–SEL 62 ms, HKG–LHR 191 ms.
HKG-1
2026.06.18
#
Maintenance
Feed B at SEL-1 was moved to new switchgear between 16:00 and 19:00 UTC on 18 June. Racks ran on feed A alone for two hours and eleven minutes, inside the design margin, and no generator start was required. No instance saw an interruption. The matching transfer on feed A is not yet scheduled; it will be posted here first.
SEL-1
2026.06.05
#
Product
Every VPS, VDS, and dedicated chassis now receives a routed IPv6 /64 at creation, at no charge, at all six metros. Existing instances can claim one from the console without a reboot. Reverse DNS is editable per address. IPv4 stays at one address per instance, with more available on request and a stated reason, as the registry requires.
All six
2026.05.22
#
Product
The Tokyo GPU row grows by twenty-four L40S cards, taking on-demand tenancy to forty cards for a single account. The rate is unchanged at $1.40 an hour, billed hourly with no reservation and no minimum. H100 stays at $2.90. Hong Kong is unchanged at $1.55 and $3.10. Your weights stay on your volume and are wiped on release.
HND-1
2026.05.09
#
Network
The transcontinental span now runs 4×100G, matching the rest of the ring. The median JFK–LAX round-trip is unchanged at 62 ms — this buys headroom, not speed, and clears the one span that ran above 60% at peak. Loss of any single span still routes the long way round without losing capacity.
JFK-1 · LAX-1
2026.04.27
#
Security
The sub-processor list is now published and versioned: every third party that touches customer data, what it does, and where it holds it. Our own controls are documented against the ISO 27001 Annex A structure — we are not certified against it, and we will say so here on the day that changes. Additions are announced here thirty days before they take effect, which is the same notice the DPA gives you to object to one.
All six
2026.04.14
#
Incident
A storage node at LAX-1 lost both paths to its NVMe shelf at 09:41 UTC, after a firmware fault in the shelf expander. Forty-six VPS went read-only rather than risk corruption; the node fenced itself and replicas took over at 09:53. No data was lost. That firmware is now pinned to the previous release at every metro. Credits were applied without a ticket.
LAX-1
2026.04.02
#
Promotion
Transfer a domain to Yunagi and the first one costs nothing beyond the registry's own year, which is added to your existing expiry rather than replacing it. DNS records, DNSSEC, and glue are carried across before the transfer is submitted, so the name never resolves to an empty zone. Portfolios above fifty names are moved by hand, over a weekend you pick.
All six
2026.03.19
#
Product
v2 of the compute API leaves beta with idempotency keys, cursor paging, and per-metro rate limits printed in the docs. The Terraform provider reaches 1.0 and covers instances, volumes, firewalls, and DNS. Median VPS deployment across the six metros is 38 seconds, measured from the API call to the first SSH banner. v1 is supported until March 2027.
All six
2026.03.05
#
Network
LHR-1 now peers at LONAP as well as LINX, taking London to 2×100G of public peering. About 9% of our UK-bound traffic moved off transit. Median LHR–JFK is 71 ms and LHR–HKG 191 ms, both unchanged: the gain is path diversity while a peer is in maintenance, not latency, and we would rather say so.
LHR-1
2026.02.20
#
Maintenance
status.yunagi.cloud is now served by a third party on infrastructure sharing nothing with AS207214 — no DNS, no transit, no power, no people on call. A status page that goes down with the platform is decoration. Ninety days of per-metro history is kept there, and the same history draws the uptime strips on the network page.
All six
2026.02.09
#
Product
Domain-validated certificates are $6 a year for a single name and $48 for a wildcard, issued in minutes against a domain you hold with us, with the validation record written for you. Renewal is automatic sixty days out unless you switch it off. If your stack already speaks ACME, free automated certificates remain the better choice, and we will say so.
All six
2026.02.02
#
Network
Our sixth metro opens in a Gasan facility with VPS only, peering at KINX and joining the private ring between Tokyo and Hong Kong. HND–SEL is 34 ms, SEL–HKG 62 ms, LAX–SEL 112 ms, LHR–SEL 249 ms. VDS and dedicated are held back until the second cage lands, which is stated here rather than implied by an empty dropdown.
SEL-1
No notices carry that tag.
過失 When We Are At Fault
One incident deserves more than a row. On 11 July an automated prefix-list rebuild withdrew HKG-1 from three of our four transit partners. The timeline below is the one kept during the event, not one reconstructed afterwards. We publish these whether or not anyone noticed and whether or not the SLA was breached, with the cause named rather than described as an unforeseen condition. Credits are applied without a ticket. If a postmortem here ever reads as though nothing could have been done differently, it is a bad postmortem, and we would rather be told so.
| time (UTC) | event |
|---|---|
| 14:02 | Prefix-list rebuild deploys to all four HKG edges at once. |
| 14:06 | Three of four transit partners stop accepting the HKG-1 /22. |
| 14:09 | Off-network probes alarm. Ring probes stay green throughout. |
| 14:11 | First customer ticket. Status page set to degraded, HKG-1 only. |
| 14:14 | Withdrawal confirmed in two external looking glasses. |
| 14:22 | Cause found: rebuild read an inventory export four days stale. |
| 14:29 | Previous prefix list re-applied at the first edge; paths return. |
| 14:37 | All four edges restored. Inbound paths fully recovered. |
| 15:10 | Status page cleared. Timeline published in this archive. |
| 12 Jul | SLA credits applied to 1,184 accounts, no ticket required. |
Sits on the ink band as a plain .tbl — page.css already restyles thead, rules and .num for .on-ink. Times are the operator log, unedited; the 12 Jul row is the only one added after the fact and is dated rather than timed.
予定 Scheduled Work
Routine work is announced fourteen days ahead, emergency work as soon as we know — twice in seven months, both inside the hour. A window is a promise about the maximum, not an estimate: if we say six hours and finish in forty minutes, the entry is updated with the time actually taken. VPS and VDS are live-migrated, so the visible effect is a pause under two seconds. Dedicated chassis are never rebooted by us; you take the reboot when it suits you, within a fortnight. Nothing in a window changes what you are billed.
curl -s https://yunagi.cloud/notices.json | jq '.entries[0]'
{
"id": "2026-09-06-hypervisor-roll",
"date": "2026-09-06",
"tag": "maintenance",
"state": "scheduled",
"sites": ["JFK-1", "LAX-1"],
"title": "Hypervisor kernel roll, JFK-1 & LAX-1",
"window": { "start": "2026-09-06T03:00:00Z",
"end": "2026-09-06T09:00:00Z" },
"impact": "live-migration pause under 2 s",
"url": "https://yunagi.cloud/notices.html#2026-09-06-hypervisor-roll"
}
# Atom https://yunagi.cloud/notices.xml
# One tag https://yunagi.cloud/notices.json?tag=incident
# Windows https://yunagi.cloud/windows/hnd-1.ics
配信 Feeds
The archive is served three ways beside the page: an Atom feed at notices.xml, the same entries as JSON at notices.json, and a webhook that fires when an entry is published or a scheduled window moves. All three carry identical fields, including the anchor, so a bot links back to exactly what a person reads. Filtering by tag works as a query string on all of them. Maintenance windows are also published as an iCalendar feed per metro, which is the honest way to get them into a change calendar rather than an inbox.
POST https://your-endpoint.example/yunagi
X-Yunagi-Event: notice.published
X-Yunagi-Signature: sha256=<hmac of the raw body>
{
"event": "notice.published",
"id": "2026-07-11-hkg-transit-withdrawal",
"tag": "incident",
"state": "resolved",
"sites": ["HKG-1"],
"opened": "2026-07-11T14:09:00Z",
"closed": "2026-07-11T14:37:00Z",
"affects": { "instances": 0, "accounts": 1184 },
"credit": "automatic",
"url": "https://yunagi.cloud/notices.html#2026-07-11-hkg-transit-withdrawal"
}
約定 What We Publish
一01
An incident that affected one harbour for thirty-one minutes is published in the same place and the same shape as a price change. Nothing is quietly left off.
二02
A maintenance window is announced here at least seven days ahead, with the harbour, the shape of the impact, and a way to move first.
三03
A notice is never deleted or rewritten. A correction is a new notice that references the old one, and the old one keeps its anchor.
問 Questions
If your question is not here, [email protected] reaches a person.
Per-instance windows appear in the console seven days ahead and are emailed to the account's technical contact. This archive is the estate-wide view; the console is the view of your own machines. Both draw from the same record, so a window cannot appear in one and not the other.
No. A single failed chassis, a single abuse suspension, or a migration you asked for is a ticket, not a notice. The line is whether an unrelated customer could have been affected: if the answer is yes, or if we are not certain it is no, it is posted here.
This is the written record; status.yunagi.cloud is the live one, refreshed every thirty seconds and hosted off our network. During an incident, watch the status page. Afterwards, read the entry here — it carries the timeline, the cause, and what changed, which a green tick cannot.
Only to add. A scheduled window gains its actual duration when the work finishes, and an incident gains its postmortem when it is written. Corrections are appended under the original text with their own date; the original stays. Nothing is silently reworded and nothing is deleted.
Every change to price, capacity or availability lands on this page first, on the day it happens.