Changes to delete behaviour with conversations

Over the past week we have changed how deletion works in our new modern interface. In this blog post I will explain what those changes were, and some actions we have taken to ensure no emails are accidentally lost.

This is a technical blog post, so it contains a moderate level of technical detail.

I will address how our backup and disaster recovery system works, and how we used it this week to recover emails which we suspected to be accidentally deleted.

Some background

Last week we rolled out the new conversations-enabled interface. However, we discovered we had under-estimated the impact of conversations on users’ existing workflow.

In particular, many users did not realise that when they selected a single item in a folder, it represented the entire conversation (all related messages, including those the user had sent).

When they pressed ‘Delete’ with one or more conversations selected, it deleted all messages in those conversations, including messages in other folders. For example, deleting a conversation in Inbox could also delete messages from “Sent Items” and “Important – Keep”.

We have altered the ‘Delete’ action to be safer in these ways:

  1. in a folder: only messages which in the same folder are deleted from the selected conversations.
  2. when viewing search results: only messages which match the search query are deleted, messages which are in the conversations but outside the search are not.
  3. when an action will cause more than one message to be deleted from a conversation, a warning message is shown to describe what will happen. The user must explicitly disable this warning if they don’t want to see it again.

What about the time before these changes?

Rather than leaving users to hunt for which emails were affected, We wrote a tool to data-mine our mail server logs. We log every create and delete of emails, along with enough data to identify which ones were “Delete to Trash”. We can also identify if the action came from an IMAP client or the web interface.

We found emails which could have been accidentally deleted using the following algorithm:

  1. the action came from the web interface.
  2. more than one message from the same conversation was deleted within 10 seconds.

All the emails which matched these criteria were restored back into the folders they were originally deleted from, with a custom keyword added. This makes it easy for users to find them again. Every affected user has been emailed with instructions on how to identify the restored emails.

How we restored data

When you delete an email on the FastMail servers, it isn’t immediately removed from disk, even if you manually expunge via IMAP. We do this:

  1. to guarantee that our “Restore from backup” feature can always find all your emails, even if they were delivered and then deleted in between backup runs.
  2. to make deletes appear faster to users.
  3. to reduce the load on our IMAP servers. Removing files is actually one of the slowest operations you can run on a modern filesystem.

So we actually batch up all deletes and run them once per week at the least busy time for our servers – Saturday night in the USA. It’s weekend everywhere in the world then.

We also never remove email files within one week after deletion, so that our “Restore” feature can work as advertised.

This is, of course, in addition to the safety provided by replication to an offsite datacentre, and daily backups to a different server running a different operating system.

Immediate response

As soon as we realised we may have to restore emails, we disabled the automated weekend cleanup job, and started collecting data from our servers.

Discoverability

The problem is that it is hard to know that an email is not there unless you actively look for it. We could disable cleanup temporarily, but not forever. Our turnover is about 2% of total email volume per week, so the disks would fill up if we never deleted anything ever again.

We decided the safe way forwards was to undo every deletion which had even the slightest chance of being by accident.

That way, if no action is taken, a few extra emails sit on disk gathering dust. It’s possible at any later time to discover them and clean them up. There is no requirement to act quickly.

We take your privacy very seriously. No contents of emails were accessed during this task, and each user’s account was processed separately to ensure there was no risk of disclosing data. You can read more about our privacy policy here: https://www.fastmail.fm/help/overview_privacy.html.

Data collection

We log every single time a message is added to or expunged from any folder on out backend servers. We collected an initial dataset of nearly 30 million “Delete to Trash” events from the log files.

The next step was identifying which of these were a single action involving more than one message from the same conversation. Every message was tagged with a session identifier and timestamp as well as the folder and IMAP “UID” which uniquely identifies it, but we were not logging the CID (conversation identifier). We do now, but that doesn’t help with log lines from the past!

Finding the CID involved writing custom code to read the index file on disk (which still contains the deleted record) and extract the CID field for every deleted message.

Finally of course, there was processing the logs for every single connection from the web servers over that time frame and finding which deletes were related to each other. There’s nothing in the log to show that it’s the same command, so we applied a heuristic of “within 10 seconds” to account for the outside case of a busy server and large folders being processed.

Restoring messages

We use the Cyrus IMAP server. One of the utilities included is called ‘unexpunge’, and it can be used to recover deleted emails. This is different from our usual restore command, which extracts messages from various sources and appends them a new temporary folder.

In this case we want to restore messages permanently, so unexpunge is the right tool… except – we want to tag every message with a keyword, and we want it to be reliable. Finding the messages afterwards is messy. We chose to add a new feature to unexpunge, setting a user-defined keyword on each message as it is restored. It is robust, and there’s no gap where messages appear without the keyword

The chosen keyword is RESTORED-20121107. Our web interface already supports global keyword search with “flag:$name”, so the email to users includes a pre-generated URL which will perform a global search on all that user’s folders for messages which were restored.

Restores are in progress now. Once they are completed, thousands of users will have some messages restored. This is almost certain to include messages which were intended to be deleted, but we must err on the side of safety here.

We have built a very robust infrastructure because of our strong commitment to data safety. These restores are in line with this commitment. It is easier to delete unwanted messages again than to recover messages which no longer exist.

Posted in Technical. Comments Off

New login and session management code on beta.fastmail.fm

We’ve just rolled out some new code on our beta server that significantly changes how sessions are managed. This new code reduces some overall session complexity, fixes some long term bugs, and adds some useful new features.

  1. There’s now just 2 main types of sessions: normal & long term
    • normal – these expire after 2 hours of inactivity
    • long term (you check the “Keep me logged in” checkbox on login) – these expire after 30 days of inactivity, for most people on most machines, this is effectively forever

    (Note: The "Keep me logged in" checkbox has been broken for the last few months on the beta server, but now correctly creates a long term login session. Also the "lightbox" login screen within the new UI now correctly works.)

  2. Logout will explicitly end a session 

    If you want to explicitly end a session, use the "Log out" link at the top right of the page. If you want to keep a session, just close the browser tab/window and when you go back to the beta server, you’ll still be logged in (see below).

  3. You can still log in to multiple different accounts

    We still support the ability to log in to multiple different user accounts at the same time on the same device/browser.

  4. You can access existing logged in sessions from the login screen

    If your device/browser has any existing logged in sessions, we now show those sessions when you go to the login screen. Simply clicking on one of those sessions will send you straight back to that mailbox for that user.

    Although by default the login screen shows existing logged in sessions, clicking the "Log in to another account" link will allow you to log in to another account at the same time.

  5. You can see (and remotely log out) all logged in web sessions on all devices/browsers

    We now track all sessions in our database and allow users to see all these sessions and remotely log out any of them individually.

    Just go to Options/Accounts –> Logged In Sessions to see all sessions in all devices/browsers. Currently only sessions created on http://beta.fastmail.fm can be deleted.

    (Note: Only web sessions are shown. IMAP/POP/XMPP/etc logins are shown on the Options -> Login Log screen)

One observation that some people might make is that with the old system, if you were logged into your account, and then closed your browser window/tab or went to http://beta.fastmail.fm again, it would appear that your existing session was automatically logged out, a nice security feature.

In fact that was never the case, the session was not logged out. Simply picking the right URL from your browser history would take you straight back in. There was just no visual indication on the login screen that this existing session was still present in your browser cookies, which is actually quite dangerous. The new system correctly shows any existing sessions on the login screen. If you want to end a session, you must use the "Log out" link at the top right of the page, whether you’re using the new system or the current system still at http://www.fastmail.fm.

Posted in News, Technical. Comments Off

Goodbye old.fastmail.fm

Summary

In early 2009 we rolled out an updated web interface to all users. This is the interface you currently see when you login at http://www.fastmail.fm as most users do.

To give users time to transition, we continued to let people login to the old pre-2009 interface if they wanted to by going to the web address http://old.fastmail.fm. We’ve continued to support this for the last 3 years, but as only a few users are still using this interface, we decided to shut it down. For the last 3 months there’s been a prominent message each time you logged into http://old.fastmail.fm that noted this, and we’ve now fully shut down http://old.fastmail.fm.

Important point: This only affects users that were explicitly going to http://old.fastmail.fm to login. Users that use the regular interface at http://www.fastmail.fm (the vast majority) are completely unaffected.

The description below is a detailed history of the old interface and includes technical details about how much things have changed since 2009 and why maintaining http://old.fastmail.fm is no longer feasible.

Goodbye old.fastmail.fm

It’s been a long road, but the old FastMail web interface has finally reached the end of its life.

You can always access your email at https://www.fastmail.fm/ or try our beta site at https://beta.fastmail.fm/.

If you want to stop reading here, the things you need to know are security concerns and it was about to break anyway. Two good reasons why now is the right time to shut the old interface down.

If you want to know some of the technical background and the technologies that we have moved through over the years, read on!

A new infrastructure

Looking back through our version control history, my very first commit was on 2004-09-20! The original web interface commits are from early 2000, though it was started before then.

We switched version control systems at some point during 2005 from CVS to Subversion, which made branching much easier – but imported all our history, so we can still look back at those early changes.

One of our major branches was a huge infrastructure switch from Redhat 7.3 to Debian 3.1 (sarge), which we worked on throughout the second half of 2005. This was all merged back into the main branch, and we converted everything over in early 2006.

http://blog.fastmail.fm/2006/01/02/one-web-server-now-running-new-infrastructure-code/

We upgraded to Debian 4.0 (etch) during May 2007, soon after it came out.

A new interface

In 2008, Neil Jenkins (who is so awesome Opera hired him even before they had decided if they were going to buy FastMail) worked as a contractor over the summer to design a more modern web interface which would take advantage of the new features in web browsers.

We branched the code, and it diverged quite considerably. Features like cross folder searching required major internal datastructure changes, and the new interface had hooks all through the code. Our plan was always to retire the old code eventually.

We released the new interface to beta at the end of 2008, and rolled it out to everyone in 2009.

http://blog.fastmail.fm/2008/11/27/help-beta-test-new-web-interface/

http://blog.fastmail.fm/2009/02/17/new-interface-being-rolled-out/

An incompatible upgrade

Then in 2009, Debian 5.0 (lenny) came out. Lenny shipped with apache2 and mod_perl2, and no longer supported apache 1.3 or mod_perl version 1. We put quite a lot of work into porting our codebase forwards to apache2. Since "old" was going away soon, we didn’t duplicate the work there.

So we installed the new web servers on lenny, and kept a couple of servers called "oldweb" still running etch. It’s amusing now to remember all the hoops I jumped through to allow automatic installation of either system.

About this time we also had machines with enough memory that 32 bit address spaces were wasteful, particularly on the IMAP servers. We moved to running 64 bit kernels with 32 bit userland.

New hardware

In 2010, the Opera sale happened. One of the early steps was to replace some of our aging hardware with equipment that was better understood and supported by the Opera sysadmin department. This meant a new bladecentre for the non-storage systems (including web)

For a little while I had two blades (redundancy!) running "oldweb" code. That’s a huge amount of very under-utilised resource.

And, to be honest, managing new blades with ancient OS was a pain. Things didn’t work well. The configuration tools we built for the new hardware didn’t run on etch.

When we moved to Debian 6.0 (squeeze) and at the same time went fully 64 bit, it was time to do something about "old".

We also moved version control systems AGAIN in late 2010 – from subversion to git. The old web servers were left on subversion, because they weren’t getting much in the way of changes any more. One more "split" in how things were done.

Fully virtual

Rather than having to support "real hardware", I built an etch virtual machine. Everything else was running squeeze 64 bit, but we still had a full 32 bit etch install path just to support oldweb.

While all this was happening, there were occasional changes required to support changing database schemas, configuration mechanisms, and interaction with other parts of our system. At some point I just took a snapshot of the current tree and started a new git repository so we could archive the subversion server entirely.

Maintaining the virtual machines was a real pain though. They were run in the background on some of the web servers to free up the hardware for more demanding tasks. This meant changing the network interfaces to be bonding drivers, custom configuration, lots of pain. There were occasionally long outages as we changed things and then had to patch oldweb to catch up.

Worst of all, we were maintaining the ENTIRE stack – support daemons, log rotation, pop fetching…

Old lost features over time – we just couldn’t keep them working, so we ripped the code out. Particularly some of the more advanced configuration screens – and everything related to billing.

Single component

In the end the virtual machines were too much work. Our authentication system in particular had many changes under the hood, and it just wasn’t going to keep working. We had a couple of really bad problems with file storage, where we were sure that something "couldn’t happen", but then it turned out old was still doing things differently. Talking to the wrong databases, running the wrong queries. We seriously considered dropping old at that point, but I wanted to give it a bit longer.

So I build a chroot installation of etch on our web servers, and bind mounted the daemon sockets into the chroot. This allowed us to run just the web interface code itself on the old branch, while running everything else in the modern, managed, outside world. I built a custom init script which could set up all the necessary mountpoints (/proc, /dev, /var/run, even the tmpfs with mmaped caches was shared) – and forward ported more of the code.

This was built with debootstrap originally, but in the end it was getting unreliable even fetching etch packages, so I build a .tar.gz file with the filesystem for the chroot, and a fresh install just unpacked that. As we changed internal config systems, I kept "oldweb" up to date. A couple of commits every month.

So that’s brings us to today. An init script (apache-oldweb), a chroot environment with a snapshot of a Debian etch machine with apache 1.3 and mod_perl version 1 – running perl 5.8. Everything else is perl 5.10 or newer, so I even have to backport some idioms as I bring back the bits which it just can’t live without.

I have done basically all the "keeping old alive" for the past couple of years – for a smaller and smaller set of users who still log in there. Backporting everyone else’s changes as they impacted old.

And etch doesn’t have security support. Hasn’t for ages. Sure it’s in a chroot, but it still has access to everything.

The final straw

But there’s one thing which oldweb can’t survive. We are redesigning how our session management works. There are some great benefits – bookmarkable URLs, remote logout of stale sessions, reduction of password typing on annoying little smartphone keyboards.

Everything will change, and old would have just stopped working. It’s not worth the changes to make it work. Particularly with the larger gap between the two systems as time goes on.

Also, and even worse, old interface is exposed to the wider internet – and it has full read/write access to the database and all emails. If there are security problems, all our users are at risk – not only those who use it directly.

It’s no longer safe, and it was going to break beyond easy repair in a few days anyway. It is time.

Goodbye oldweb

Posted in News, Technical. Comments Off

One step forward, two steps back

It’s been a really bad week for me.  Backing out two significant pieces of work.  One only released recently, but the other having caused problems for an entire year, and I’m really sorry to those who’ve been sitting through them as we didn’t have the effort available to find the underlying cause.

PowerDNS

First the recent one.  We switched out DNS servers from tinydns to powerdns last week.  There were very good reasons for the switch, tinydns as-is doesn’t support IPv6, or DNSSEC, or zone transfers, or…

And the data file is built as a single giant database and synchronised to all our servers once per hour, so updates take some time to be made.

On the flip side – it’s rock solid!  It’s served us well for years.  So we put a lot of work into testing PowerDNS for the change.  Unfortunately, it wasn’t enough.  First SOA records were broken for subdomains, then DNS delgation didn’t work, and now that I’ve switched back, a problem with Chat server aggregation has gone away, so it was probably doing the wrong thing there too!

Anyway – powerdns got backed out.  The “pipe” backend that we were using just isn’t expressive enough, so we either need to find another way to do it, or find another path forward.  The good thing about PowerDNS is that it’s actively maintained, so we should be able to get somewhere here.

EJabberd

It’s much sadder to give up EJabberd.  Erlang is an interesting language, and the integration work was done by an intern last year – he did really good work.  The hard bit was that we needed support for the many thousands of domains that our customers host with us.  Ejabberd 2.x (the stable branch) just didn’t support it.  Ejabberd 3.x was going to, but was currently in alpha.  Looking at the development pace, I made the call to integrate with Ejabberd.  We did that.

But it’s been plagued with problems.  The chat logging service has been flaky, there have been “Malformed XML” disconnections which I suspect to be related to incorrect SSL renegotiations, but I haven’t been able to prove it.  I’ve spent far too much time looking at packet logs and trying to figure it out.

I’ve had long standing tickets about it, and kept saying “it’s getting better” – but seriously, upstream hasn’t made a single commit to ejabberd mainline since February this year.  They’re putting all their effort into the 2.x branch.

So I’m in the process of backporting our chat service to the DJabberd engine we used to use.  It’s not perfect either – it doesn’t have anywhere near the feature set that ejabberd has, and it’s not getting any more support.  The code is of OK quality, but it’s quite convoluted and written in many different styles which makes reading it tricky.  I’ve had to make two patches to get interoperability up to scratch with modern servers and support the multiple SSL certificates we now use.

It’s always sad to give up features, and to sideline hard work that you or others have done – but in the end we have been hurting customers by providing a sub-standard experience with chat.  So I’m hoping to put a line under that by the end of this week and be able to move on with good new things again.  At least a couple of us have some more Erlang experience now, and you never know when that might be useful.  It’s good just to understand different ways of thinking about code.

Posted in Technical. Comments Off

*.fastmail.fm certificate updated

The SSL certifcate for *.fastmail.fm (that is, http://www.fastmail.fm, and all other subdomains) was due to expire in a couple of months, so this morning we’ve updated it to a new one.

Because its not something we do often, we don’t have a nice little script automating the task. As a result I forgot to properly add the CA chain to the certificate so there was a period of about 30 minutes where some users with old browsers might have seen “invalid certificate” errors. That was fixed pretty quickly once we noticed and everything should be fine now.

The next certificate to expire has a few months left on it. I’ll make sure this is nicely automated before we have to update that one!

Posted in Technical. Comments Off

Chat server updated

We’ve had a couple of persistent problems with our XMPP chat server since upgrading to Ejabberd (3.0 pre-release with patches) last year.

  1. “Malformed XML” errors.  We suspect these were due to bugs with handling SSL.
  2. Missing IP address and SSL information in the LoginLog.
  3. ChatLog backend died, and chat logging stopped.

We have our awesome intern Samuel back again this year, and though he’s working with a different department this time – I managed to grab some of his time to work with me on fixing these long standing issues!

We think we’ve fixed number 1 – at least it hasn’t been seen since updating the upstream libraries.

We’re positive we’ve fixed number 2, by making sure the correct information is passed directly to the authentication functions rather than looking for it in the session object (which isn’t populated in time to be useful)

And as for number 3 – a restart fixed it.  I’m now monitoring the log files to look for what might cause it to die.  Unfortunately, by the time we looked at it the logs had already rotated away, so there was no evidence left!  I’ll make sure we catch the problem next time, if it happens again.
TL;DR – chat server is updated and should be more stable and give better information than before.

 

Posted in Technical. Comments Off

A story of leaping seconds

I’ve been promising to blog more about some technical things about the FastMail/Opera infrastructure, and the recent leap second fiasco is a good place to point out not only some failures, but also the great things about our system which helped limit the damage.

Being on duty

First of all, it was somewhat random that I was on duty. We’ve recently switched to a more explicit “one person on primary duty” from the previous timezone and time-of-day based rules. Partially so other people knew who to contact, and partially to bring some predictability to when you had to be responsible.

Friday 29th June was my last working day for nearly 3 weeks. My wife is on the other side of the world, and the kids are off school. I was supposed to be handing duty off to one of the Robs in Australia (we have two Robs on the roster – one Rob M and the other Rob N. This doesn’t cause any confusion because we’re really good at spelling). I had made some last minute changes (as you always do before going away) and wanted to keep an eye on them, so I decided to stay on duty.

Still, it was the goodbye party for a colleague that night, so I took the kids along to Friday Beer (one of the awesome things about working for Opera) and had quite a few drinks. Not really drunk, but not 100% sober either. Luckily, our systems take this into account… in theory.

The first crash

Fast forward to 3:30am, and I was sound asleep when two SMSes came in quick succession. Bleary eyed, I stumbled out to the laptop to discover that one of our compute servers had frozen. Nothing on the console, no network connectivity, no response to keystrokes via the KVM. These machines are in New York, and I’m in Oslo, so kicking it in person is out of the question – and techs onsite aren’t going to get anything I don’t.

These are Dell M610 blades. Pretty chunky boxes. We have 6, split between two bladecentres. We can easily handle running on 3, mail delivery just slows by a few milliseconds – so we have each one paired with one in another bladecentre and two IP addresses shared between them. In normal mode, there’s one IP address per compute server.  (amongst the benefits of this setup – we can take down an entire bladecentre for maintenance in about 20 minutes – with no user visible outage)

Email delivery is hashed between those 6 machines so that the email for the same user goes to the same machine, giving much better caching of bayes databases and other per-user state. Everything gets pushed back to central servers on change as well, so we can afford to lose a compute server, no big deal.

But – the tool for moving those IP addresses, “utils/FailOver.pl”, didn’t support the ‘-K servername’ parameter, which means “assume that server is dead, and start up the services as if they had been moved off there”.  It still tries to do a clean shutdow, but on failure it just keeps going. The idea is to have a tool which can be used while asleep, or drunk, or both…

The older “utils/FailOverCyrus.pl”, which is still used just for the backend mail servers, does support -K. It’s been needed before. The eventual goal is to merge these two tools into one, which supports everything. But other things keep being higher priority.

So – no tool. I read the code enough to remind myself how it actually works, and hand wrote SQL statements to perform the necessary magic on the database so that the IP could be started on the correct host.

I could have just bound the IP by hand as well. But I wanted to not confuse anything. Meanwhile I had rebooted the host, and everything looked fine, so I switched back to normal mode.

Then (still at 3:something am) I wrote -K into FailOver.pl. Took a couple of attempts and some testing to get it right, but it was working well before I wrote up the IncidentReport and went back to bed some time after 4.

The IncidentReport is very important – it tells all the OTHER admins what went wrong and what you did to fix it, plus anything interesting you saw along the way. It’s super useful if it’s not you dealing with this issue next time. It included the usage instructions for the new ‘-K’ option.

Incident 2 – another compute server and an imap server

But it was me again – 5:30am, the next blade over failed. Same problem, same no way to figure out what was wrong. At least this time I had a nice tool.

Being 5:30am, I just rebooted that blade and the other one with the same role in the same datacentre, figuring that the thing they had in common was bladecentre2 compute servers.

But while this was happening, I also lost an imap server. That’s a lot more serious. It causes user visible downtime – so it gets blogged.

From the imap server crash, I also got a screen capture, because blanking wasn’t enabled for the VGA attached KVM unit. Old server. This capture potentially implicated ntpd, but at the time I figured it was still well in advance of the leap second, so probably just a coincidence. I wasn’t even sure it was related, since the imap server was an IBM x3550, and the other incidents had been blades. Plus it was still the middle of the night and I was dopey.

Incident 3

Fast forward to 8:30am in the middle of breakfast with my very forgiving kids who had let me sleep in. Another blade, this time in the other blade centre. I took a bit longer over this one – decided the lack of kernel dumps was stupid, and I would get kdump set up. I already had all the bits in place from earlier experiments, so I configured up kdump and rolled it onto all our compute servers. Rob M (see, told you I could spell) came online during this, and we chatted about it and agreed that kdump was a good next step.

I also made sure every machine was running the latest 3.2.21 kernel build.

Then we sat, and waited, and waited. Watched pots never boil.  I also told our discussion forum what was going on, because some of our more technical users are interested.

Incident 4

One of the other IMAP servers, this time in Iceland!

We have a complete live-spare datacentre in Iceland. Eventually it will be a fully operational centre in its own right, but for now it’s running almost 100% in replica mode. There are a handful of internal users with email hosted there to test thing (indeed, back at the start, my “urgent” change which meant I was still on call was changing how we replicate filestorage backends around so that adding a second copy in Iceland to be production rather than backup didn’t lead to vast amounts of data being copied through the VPN to New York and back while I was away)

So this was a Dell R510 – one of our latest “IMAP toasters”. 12 x 2Tb SATA drives for data (in 5 x RAID1 and 2 hot spares) for 20 x 500Gb cyrus data stores, 2 x SSD in RAID1 for super-hot Cyrus metadata, 2 x quadcore processor, 48 Gb RAM. For about US $12k, these babies can run hundreds of thousands of users comfortably. They’re our current sweet spot for price/performance.

No console of course, and no kdump either. I hadn’t set up the Iceland servers with kdump support yet. Doh.

One thing I did do was create an alias. Most of our really commonly used commands have super short aliases. utils/FailOver.pl is now “fo”. There are plenty of good reasons for this.

Incident 5

One reason is phones. This was Saturday June 30th, and I’m taking the kids to a friend’s cabin for a few days during the holidays. They needed new bathers, and I had promised to take them shopping. Knowing it could be hours until the next crash, I set up the short alias, checked I could use it from my phone, and off we went. I use Port Knocker + Connect Bot from my Android phone to get a secure connection into our production systems while I’m on the move.

So incident 5 was another blade failing – about 5:30pm Oslo time, while I was in the middle of a game shop with the kids. Great I thought, kdump will get it. I ran the “fo” command from my phone, 20 seconds of hunt and peck on a touch keyboard, and waited for the reboot. No reboot. Couldn’t ping it.

Came home and checked – it was frozen. Kdump hadn’t triggered. Great.

As I cooked dinner, I chatted with the sysadmins from other Opera deparements on our internal IRC channel. They confirmed similar failures all around the world in our other datacentres. Everything from unmodified Squeeze kernels through to our 3.2.21 kernel. I would have been running something more recent, but the bnx2 driver has been broken with incorrect firmware since then – and guess what the Dell blades contain. Other brands of server crashed too, so it wasn’t just Dell.

Finding a solution

The first thing was to disable console blanking on all our machines. I’ve been wanting to do it for a while, but I down-prioritised it after getting kdump working (so I thought). By now we were really suspicious of ntp and the leap second, but didn’t have “proof”. A screen capture of another crash with ntp listed would be enough for that. I did that everywhere – and then didn’t get another crash!

Another thing I did was post a question to superuser.com. Wrong site – thankfully someone moved it to serverfault.com where it belonged. My fault, I have been aware of these sites for a while, but not really participated. The discussion took a while to start up, but by the time the kids were asleep, it had exploded. Lots of people confirming the issue, and looking for workarounds.  I updated it multiple times as I had more information from other sysadmins and from my own investigations.

Our own Marco had blogged about solutions to the leap second, including smearing, similar to what Google have done.  I admit I didn’t give his warnings the level of attention I should have – not that any of us expected what happened, but he was at least aware of the risk more than some of us cowboys.

Marco was also on IRC, and talked me though using the Perl code from his blog to clear the leapsecond flag from the kernel using adjtimex. I did that, and also prepared the script a bit for more general use and uploaded it to my filestorage space to link into the serverfault question where it could be useful to others.

By now I was the top trending piece of tech news, and scoring badges and reputation points at a crazy rate. I had killed NTP everywhere on our servers and cleared the flag, so they were marching on oblivious to the leap second – no worries there. So I joined various IRC channels and forums and talked about the issue.

I did stay up until 2am Oslo time to watch the leap second in.  In the end the only thing to die was my laptop’s VPN connection, because of course I hadn’t actually fixed the local end (also running Linux).  There was a moment of excitement before I reconnected and confirmed that our servers were all fine.  10 minutes later, I restarted NTP and went to bed.

The aftermath: corruption

One of the compute server crashes had corrupted a spamassassin data file, enough that spam scanning was broken.  It took users reports for us to be aware of it.  We have now added a run of ‘spamassassin –lint’ in startup scripts of our compute servers, so we can’t operate in this broken state again.

We also reinstalled the server.  We reinstall at almost any excuse.  The whole process is totally automated.  The entire set of commands was

# fo -a    (remember the alias from earlier)
# srv all stop
# utils/ReinstallGrub2.pl -r     (this one has no alias)
# reboot

and about 30 minutes later, when I remembered to go check

# c4 (alias for ssh compute4)
# srv all start
# c1 (the pair machine for compute4 is compute1)
# fo -m (-a is “all”, -m is “mastered elsewhere”)

And we were fully back in production with that server.  The ability to fail services off quickly, and reinstall back to production state from bare metal, is a cornerstone of good administration in my opinion.  It’s why we’ve managed to run such a successful service with less than one person’s time dedicated to sysadmin.  Until 2 months ago, the primary sysadmin has been me – and I’ve also rewritten large parts of Cyrus during that time, and worked on other parts of our site too.

The other issue was imap3 – the imap server which crashed right back during incident 2.  After it was up, I failed all Cyrus instances off, so it’s only running replicas right now.  But the backup script goes to the primary location by default.

I saw two backup error messages go by today (while eating my breakfast – so much for being on holiday – errors get sent to our IRC channel and I was still logged in).  They were missing file errors, which never happen.  So I investigated.

Again, we have done a LOT of work on Cyrus over the years (mostly, I have), and one thing has been adding full checksum support to all the file formats in Cyrus.  With the final transition to the twoskip internal database format I wrote earlier this year, the only remaining file that doesn’t have any checksum is the quota file – and we can regenerate that from our billing database (for limit) and the other index files on disk (for usage).

So it was just a matter of running scripts/audit_slots.pl with the right options, and waiting most of the day.  The output contains things like this:

# user.censored1.Junk Mail cyrus.index missing file 87040
# sucessfully fetched missing file from sloti20d3p2

# user.censored2 cyrus.index missing file 18630
# sucessfully fetched missing file from slots3a2p2

 The script has enough smarts to detect that files are missing, or even corrupted.  The cyrus.index file contains a sha1 of each message file as well as its size (and a crc32 of the record containing THAT data as well), so we can confirm that the file content is undamaged.  It can connect to one of the replicas using data from our “production.dat” cluster configuration file, and find the matching copy of the file – check that the sha1 on THAT end is correct, and copy it into place.

The backup system knows the internals of Cyrus well.  The new Cyrus replication protocol was built after making our backup tool, and similar checks the checksums of data it receives over the wire.  At every level, our systems are designed to detect corruption and complain loudly rather than blindly replicating the corruption to other locations.  We know that RAID1 is no protection, not with the amount of data we have.  It’s very rare, but with enough disks, very rare means a few times a month.  So paranoia is good.

Summary

All these layers of protection, redundancy, and tooling mean that with very little work, even while not totally awake, the entire impact on our users was an ~15 minute outage for the few users who were primary on imap3 (less than 5% of our userbase) – plus delays of 5-10 minutes on incoming email to 16.7% of users each time a compute server crashed.  While we didn’t do quite as well as Google, we did OK!  The spam scanning issue lasted a lot longer, but at least it’s fully fixed for the future.

We had corruption, but we have the tools to detect and recover, and replicas of all the data to repair from.

Posted in Technical. Comments Off
Follow

Get every new post delivered to your Inbox.

Join 5,591 other followers