Other logs have fallen below 90% and survived (Sectigo has a log recently that did this) so availability is not the instant removal that it used to be.
The log is still operational and none of our alarms are triggering as it is failing for monitors, not CAs or on our internal systems.
We're investigating that and will adjust the log's operation as needed. Of course, if Google wants to retire the log then that's fine as well. However, it would be good to know if it will be retired right away so we don't waste time troubleshooting and fine-tuning the logs scale.
Other logs have fallen below 90% and survived (Sectigo has a log recently that did this) so availability is not the instant removal that it used to be.This is a decidedly concerning response.Log operators commit to a policy which requires 99% availability. (Note that the 24h uptime is informational. The policy measures a 90d rolling window, in which this log currently has 99.1662% availability.) The fact that logs were allowed to violate that policy in the past was an unfortunate necessity when logs were critically few, it's not a permanent license to disregard it. As a community, we should strive to improve the health of the ecosystem, not take opportunities to ratchet it down.
The log is still operational and none of our alarms are triggering as it is failing for monitors, not CAs or on our internal systems.Claiming that the log is operational is even more puzzling: monitors are the main customers of CT logs. If monitors report being unable to keep up with the log, the log is unavailable. The reason MMDs exist is to bound the time between certificates being issued and certificates reaching monitors. If monitors are more than 24h behind, that is an ongoing outage.
We're investigating that and will adjust the log's operation as needed. Of course, if Google wants to retire the log then that's fine as well. However, it would be good to know if it will be retired right away so we don't waste time troubleshooting and fine-tuning the logs scale.This feels misplaced. Google made the requirements very clear. If you are unable or unwilling to meet them, you should retire the log, not let Google do it for you, and on top of that demand they deliver that decision urgently for your convenience.
I know this might be strident coming from a fellow CT log operator, but the whole reason Geomys runs a log is to demonstrate that it is possible to run a compliant log with a limited budget, and help fellow operators do so. If well-funded operators argue against the importance of the bare-minimum availability bar, that kind of defeats the point.
Other logs have fallen below 90% and survived (Sectigo has a log recently that did this) so availability is not the instant removal that it used to be.This is a decidedly concerning response.Log operators commit to a policy which requires 99% availability. (Note that the 24h uptime is informational. The policy measures a 90d rolling window, in which this log currently has 99.1662% availability.) The fact that logs were allowed to violate that policy in the past was an unfortunate necessity when logs were critically few, it's not a permanent license to disregard it. As a community, we should strive to improve the health of the ecosystem, not take opportunities to ratchet it down.No, it's not a concerning response. The response is pointing out that historically this 99% has not triggered an immediate removal of the log. If this was a requirement that mandated immediate removal, then we would turn down the log now. The history with logs means we aren't sure if this log is not viable or not based solely on availability to monitors.The log is still operational and none of our alarms are triggering as it is failing for monitors, not CAs or on our internal systems.Claiming that the log is operational is even more puzzling: monitors are the main customers of CT logs. If monitors report being unable to keep up with the log, the log is unavailable. The reason MMDs exist is to bound the time between certificates being issued and certificates reaching monitors. If monitors are more than 24h behind, that is an ongoing outage.That is not correct either. MMD is the mean time to merge, not the mean time to pull. It could be considered an outage I suppose, but you can't redefine MMD.
Hi folks,
We had not chimed in on this thread until now because our compliance monitoring infrastructure still has sphinx2026h1 above 99% 90-day availability, albeit barely. However, I wanted to clarify our view on the importance of log availability, and how we form our policy responses.
I would direct attention to Chrome's CT Log Policy section on rate limiting. In particular, it reads in part, "Log data availability is of paramount importance. All rate limits must be set to ensure that all well-behaved clients can reliably retrieve log entries at a rate greater than the growth rate of the log."
The policy also notes that "removal of the log from Chrome’s log list may not be the best outcome for Chrome, Chrome’s users, or the CT ecosystem." As a result, dropping below 99% availability does not automatically result in the retirement of a log by our policy.
In the CT ecosystem of a year ago, we were deeply concerned that the removal of several logs risked the viability of the overall ecosystem. We had 25% fewer log operators then, with the average log offering a much lower quality of service than logs today. Ultimately, our policy decisions, now and then, consider both the ecosystem's health, and all information available from the incident, including how the operator responds to the incident. I'd direct attention to the corresponding policy section on incident detection and response.
Thanks for working to address the issue, Jeremy. I hope we can look forward to a prompt resolution, and assuming sphinx2026h2's availability does indeed drop below 99%, a full postmortem once the technical issues have been addressed.
Best,
Joe, on behalf of the Chrome CT Team
Status right now
As mentioned, Sphinx 2026h2 write access is off as of 4:23 PM Mountain today, July 22nd. We will keep it off until until we have a fix for the read throttling and confirmation that monitors have caught up. We are looking at Wyvern to see if that log has the same issue.
What we've done so far is temporarilty remove the read throttling on Sphinx get-entries. That produced a short window where pull rates improved. After about half an hour, the log came under enough read load that it started to fail, and we had to reimpose a limit. The limit we put in place is higher than the one Andrew and Colin have been hitting. Andrew, Colin, Ben - if you could check again and let me know if you're seeing improvement then that would help. We are watching the numbers overnight ourselves and will adjust as needed, but I thought your observations would be useful in determining the read-write balance.
What caused this
Certificate volume on Sphinx 2026h2 spiked starting July 16th. Write volume was manageable. The read side was not withget-entries traffic running at roughly 50 percent higher volume than write traffic. On top of that, each read request pulls a batch of entries compared to the one certificate per write, making the read data close to 100x the write data. The rate limit on get-entries was not sized for that increase, which is why the monitors have been hitting the limits over the last few days.
Why we did not catch this sooner
This is the part Cynthia asked about directly, so I want to answer it straight rather than fold it into the technical explanation. Note this is my initial analysis and not the fulll write-up of what happened. Details will likley change as the investigation is on-going.
Rick's July 17th post described the CTile misconfiguration as resolved. We treated the elevated read latency that followed as a residual symptom of that incident and spike in logging rather than a separate, ongoing one. Monitors falling behind the logging rate is its own issue and should have been separated out from the July 16th post.
Separately, our CT mailing list coverage failed at the same time. We keep three people on rotation monitoring the list. Two were on approved personal leave and one was pulled into a personal emergency with no advance notice. The people covering for them did not realize this thread needed urgent handling, so it sat as a routine item instead of an incident. This was because none of our internal alerts were firing as the system was not considered "down". That is a gap in our on-call handoff process.
As mentioned, our internal alerts did not fire because they are built around service availability, not around read throughput relative to monitors. We have an alert for monitors but only that the read-access is working, not that the read-access speed has dropped below write-access speed. The log was considered up, even if access was too slow for monitors to keep pace.
We're still working on optimizing access but wanted to give you all an update on where we are at and the preliminar findings.
Jeremy