My Journey with ZFS on Linux

Toward the latter half of 2016, I decided to convert my home file server’s file system to ZFS as an experiment on real hardware in my own test environment. I knew before embarking on this journey that ZFS on Linux (hereafter referenced as ZoL) is stable but not production ready. The experiment is nearing its end, and I’ll be switching to an mdadm-based array with ext4 for reasons I’ll be exploring in this post. If your storage needs require the redundancy and self-healing properties of ZFS, consider other operating systems where it’s better supported and stable, like FreeBSD.

This post illustrates my thinking as of early 2017 and why I’m not going to continue with ZoL (for now).

Conclusions

I don’t wish to bore you with excess detail, so I figured I’d present my conclusions up front. You can then decide if the rest of this post is worth reading. Let’s begin exploring.

First, using ZoL in a home file server environment presents us with the following pros and cons:

Pros

  • Transparent compression can save a surprising amount of space with with LZ4 (useful!)
  • Lower administrative overhead (a wash; not always true in practice)
  • Data integrity guarantee (always make backups)
  • Self-healing (always have backups)
  • It’s not Btrfs (you’ll probably use your backups)

Cons

  • May require more extensive deployment planning
  • Some applications may require dataset tweaking
  • Administrative overhead not always as advertised (see above), but this is mostly the fault of the current state of ZoL
  • Poor performance for some workloads (especially databases; versus ext4)
  • Lack of recovery tool support limits possible restoration options when things go wrong (backups!)

Noteworthy bullet points when using ZoL:

  • Stable but not production ready (an important distinction!)
  • Upgrades will be painful!
  • /boot on ZFS is possible but…
  • DKMS ZFS modules are a horrible idea; don’t do this–don’t ever do this
  • Always create bootable media with ZFS tools installed in case something goes wrong (it will–don’t forget backups!)

The Meat and Potatoes of ZoL

I would not recommend ZoL for anything outside a test environment at this time (I’ll explain in a moment). ZoL may be useful for long term storage or a backup NAS box if you’re particularly brave. However, if you plan on deploying ZFS permanently or at least deploying it in a long term installation, I’d recommend using an OS with which ZFS is tightly integrated; recommendations include FreeBSD and FreeBSD-based distributions (including FreeNAS) or OpenIndiana-based platforms. I’d also shy away from using ZFS on systems that are intended to be used as general purpose machines; in my experience, ZFS really shines in a heavily storage-centric configuration, such as NAS, where throughput isn’t as important as integrity. Outside this environment, ZFS may be a performance liability, and current benchmarks demonstrate underwhelming behavior when used as a backing store on Linux for databases. In effort to work around this, ZFS requires planning and tuning for use with RDBMSes and may also impact write-heavy applications. Read-only loads are something of a wash–file system choice is better made with regards to workload requirements: Is raw performance or data integrity more important? Note that and the latter–data integrity–can be solved via other means, depending on use case, but ZFS’s automated self-healing capabilities are hugely attractive in this arena.

As of early 2017, ZFS on Linux is still exceedingly painful to install and maintain for rolling release distributions. In particular, on Arch Linux, ZFS will cripple most upgrade paths available for the kernel, itself, and applications that require ZFS dataset tuning. It’s almost impossible to upgrade to more recent kernels if you pride your sanity (and system stability) unless you expect to wait until such time as the latest kernel has been tested, new ZFS PKBUILDs are ready, and you’re open to an afternoon of potential headaches. Generally speaking, the upgrade process itself isn’t frustrating, but it should always be preceded by a fresh backup–and keep your rescue media handy!

Never use ZFS with DKMS in effort to shortcut kernel versioning requirements, even if upstream ZoL appears to support the new version–the package you’re using may not be updated yet, and the AUR DKMS PKGBUILDS for ZFS are not as stable or well maintained as the kernel-pinned packages. With DKMS, if there’s a mismatch, even slightly, you risk potential kernel panics and a long, painful recovery process. I was bit by this early on and discovered that some kernel panics aren’t immediate; instead, they may occur after an indeterminate period of time depending on file system activity, and the recompilation process of ZFS and SPL will consume a sufficient amount of time such that your system will likely panic before the build completes. Stick with the version-fixed builds; don’t use the DKMS modules. Of course, this introduces a new problem: Now you have to pin ZFS to a specific kernel version, leading to extra work on rolling release distributions like Arch…

It’s necessary to plan for a lot of extra work when deploying ZoL. Estimate how much time you’re likely (or willing) to spend and multiply it by two. I’m not even kidding. ZoL on distributions without completely pre-built packages (possibly others) require all of the following: Building SPL, its utilities; ZFS and its utilities; installing unstable versions of grub if you plan on keeping /boot on ZFS (with the implication that you’re one upgrade away from an unbootable system at all times); dataset tweaking per application; and potential bugs. Lots of them. When I first deployed ZFS on Linux, I was completely unaware of a deadlock condition which affected the arc_reclaim kernel process, and everytime I’d run rsync, arc_reclaim would hang, CPU usage would spike, and it’d be necessary to manually intervene. To say nothing of the RAM usage…

ZFS performance under Linux is also relatively poor for my use case. While read/write speeds are acceptable, it’s nowhere near ext4, and it’s likely slower than ZFS on the same hardware running FreeBSD. Furthermore, the memory pressure due to the ZFS driver’s ARC failing to release RAM back to the system (it’s supposed to, but I’ve yet to see it in practice) under memory-intensive operations can cause an out-of-memory condition, swapping, and a potentially fatal invocation of the Linux OOM killer. For this reason alone, I could never recommend ZoL on a general purpose server. If your deployment is intended exclusively as a NAS and you have literally no other services besides NFS/Samba, ZFS’ memory usage would be perfectly fine, provided it’s on a system with 8+ GiB RAM (though you should have more). Then there’s the possibility of L2ARC devices, caching devices, and so forth. If you’re planning on running other services in addition to a ZFS-backed NFS server, such as GitLab or Minecraft, you’ll quickly find that striking a balance between RAM allocation for the ARC versus other applications becomes a somewhat tedious chore. In fact, I might even venture as far as to suggest that you shouldn’t consider running ZFS plus other applications on anything less than 16 GiB RAM–preferably 32 to hand off a nice big chunk to the ARC, particularly if you plan on expanding drive capacity, and you still shouldn’t run anything you don’t absolutely need (seriously: Make it a pure NAS).

Tweaking ZFS for database loads doesn’t seem particularly noisome–certainly not on the surface–until you encounter one or more upgrade cycles that require more than just a dump/load. If you follow the default Arch Linux upgrade process for PostgreSQL, you’ll quickly find ZFS less flattering than alternatives. Not only is it necessary to tweak the recordset size in addition to a few other file system attributes for performance reasons (though you only do this at dataset creation), but suddenly, following the Arch upgrade guide by moving the old data directory and creating a new one is plagued with matters of shuffling around extra files, remounting previously tweaked datasets to the appropriate location, and more. In my case, I had to copy all of the original data to a new directory, wipe the old mount point for both the data directory and the write-ahead log, then create a new data directory in the old dataset mount point, copy the WAL into the WAL-specific mount point, mount it at pg_xlog, and only then could I complete the upgrade process. MySQL on ZFS is generally easier to upgrade, in my experience, but I also use it much less frequently. Be aware that MySQL still requires dataset tweaks, and the tweaks applied depend on whether you’re using primarily MyISAM or InnoDB. I’ve not experimented sufficiently to understand whether it’s possible to tweak datasets for individual storage engines.

Of course, there’s a few other relatively minor niggles that depend on your specific case. For example, grub2 naturally has no understanding of ZFS (or XFS for that matter), so it’s necessary to install grub-git if your /boot partition isn’t on a separate ext3/4 install. Under Arch, it’s also necessary to make certain your initrd is correctly configured via /etc/mkinitcpio.conf, and it’s almost always a good idea to re-run mkinitcpio even after upgrading or installing the kernel just in case it didn’t pick up your ZFS modules (you may need the binaries, too). Otherwise, you’ll be resorting to your emergency boot media to fix the dilemma (you did create it, didn’t you?).

A Less Optimal Solution

I consider the experiment with ZFS on Linux a fantastic success even though I’m already planning to migrate away from it. For my needs, I’m reluctant to run FreeBSD even though I’ve used it for similar purposes in the past. Thus, I’ll be reinstalling the machine with a combination of ext4 + mdadm (actually copying it back over, but there’s no functional difference insofar as downtime). In retrospect, I’ll probably miss ZFS’ transparent compression the most. Given my relatively modest data size and the fact that it defaults to lz4 compression (speed optimized), it’s surprising that it’s saved close to 200 GiB of storage! No other file system, save for Btrfs, provides transparent compression, and in spite of the integrity guarantees ZFS provides, I think compression is a far more pragmatic upside since its impact is real and immediate rather than theoretical.

Although I’d like to wax philosophical about ZFS’ touted benefits, I still can’t help but think it’s solving a problem that is gratuitously overblown. Perhaps bitrot is a real, material risk, but I’ve rarely been affected by it (ancient backup CDROMs notwithstanding). Has it affected my archives? Almost certainly so, but if it has, it’s never had an impact on photos or media, much less other, more critical data; the few times I’ve performed checksum validation of archives versus physical disk contents, I haven’t encountered a mismatch. Indeed, although it’s a problem ZFS is almost tailor-made to fix, it still doesn’t beat regular, extensive backups. Of course, that assumes you have a mechanism in place that would prevent your backups from being adulterated or overwritten by later, corrupted snapshots (and that your backups aren’t subject to bitrot as well), but I think Google’s solution here is far more apropos: Keep no fewer than three copies of your most important data. Surely one of them, statistically, will survive.

You’ll notice that I haven’t mentioned ZFS snapshots (or send/receive), because I’ve yet to encounter a circumstance (besides upgrades, perhaps) where they’re useful to me. While I’d like to use them with containers, there’s still the very real problem of running software inside a container that requires dataset tweaking, and there’s also the specter of lingering problems with ZoL’s implementation of ZFS which has had problems as recently as last year with snapshots (mostly with send/receive if memory serves). In my case, I tend to avoid advanced features if there’s a risk of causing damage because they’re either not well-tested, buggy, or have had a recent history of inducing failures. But alas, I’m conservative at heart; I’d happily poke about with ZFS snapshots in a virtual machine to see how they work, but I’m much less happy about using them on a real server that’s doing real work where downtime would likely interfere with important activities when those same kernel drivers are of dubious stability. I also have no other ZFS systems where send/receive would benefit me.

There is an alternative file system some of the more astute readers among you may have noticed in my list of omissions: Btrfs. I considered Btrfs for my server, testing it briefly, but at the time (mid-2016), I encountered some evidence that suggested Btrfs may not be particularly stable in spite of it being among the list of default file systems for some distributions. Btrfs’ tools feel lacking, dampening my confidence further.

The Btrfs authors have as recently as August 2016 admitted to substantial, possibly unfixable problems with then-current Btrfs’ RAID5/6 implementations. Although I’m running a simple mirror, the fact that such bug would be present in a file system some distributions have optimistically labeled as “stable” is worrisome (but just don’t use its RAID5/6 features–or whichever other features happen to be broken). I’ve seen comments from as early as 2014/2015 lauding the benefits of Btrfs as a stable, tested platform, but I have serious reservations substituting caution with optimism, particularly when 1-2 years later in 2016, it would appear such optimism is horribly misplaced. Consequently, I don’t see Btrfs as a viable alternative (yet!), and that’s without addressing Btrfs’ performance history with regards to PostgreSQL and other databases. This may change, eventually, and Btrfs does look promising. There are some valid criticisms that Btrfs is simply reinventing ZFS (poorly), but being as ZFS will likely never be included in the kernel due to licensing conflicts, Btrfs is better poised to reach parity with the likes of ext4 and company. I’m not optimistic about the Btrfs timeline, and while I’d be surprised if it attains feature completeness before 2020, I do believe many of its misgivings will be resolved well before then.

Back to ZFS: Will I revisit it in the future? Absolutely. If I do, I may put together a FreeBSD NAS for additional storage or the likes. For now, however, ext4 is suitable enough for my needs. It’s certainly good enough for Backblaze, and while I don’t share their requirements, I take a more pragmatic approach. If ext4 is good enough for them, it’s good enough for me.

No comments.
***

systemd, bridges, containers, and IPv6

I apologize for the quickie here; it’s late, I’m tired, and I just cobbled together a solution to a problem that’s been bothering me since this weekend. First, some backstory since I know you guys appreciate it when I go off on tangents. (And if you don’t, you can just scroll passed this nonsense.)

I’ve been migrating most of my services on my home file server into their own containers. I already did this on my VPS, and I plan on doing it to others eventually, because the service isolation is somewhat helpful and prevents stupid mistakes from doing equally stupid things to the host. Containers aren’t a panacea, of course, and the man pages specifically warn against using them as a security apparatus. It’s not effective, so they say, but I’m also a believer in defense-in-depth.

Anyway, the intent has been mostly a matter of isolation. I have a Minecraft server running, among other things, and I’ve been wanting to isolate it from the rest of the system. Not only does this make configuration somewhat easier since I can have service-specific user accounts without polluting the host passwd entries, but it means I can provide some package isolation as well (no Java on the host–yay!). This isn’t without its shortcomings, mind you, and it’s been one Hell of an interesting battle, particularly when IPv6 is thrown into the mix. But I digress, and the server-specific configuration is a welcome post for another day.

In the coming weeks, I’ll see about making a post on using systemd-nspawn for container isolation, why I chose it over Docker and LXC, how to get it working with user namespaces, and how to configure it with IPv6 addresses routed to each container (rather than through auto-config). Hint: you need routable prefixes like a /48 or /56–trying to (ab)use the neighbor discovery proxy and the respective sysctl settings to segment parts of your /64 non-routable prefix won’t work. Don’t even bother, because it’s not worth the misery.

Meanwhile, back to business: Since I do quite a bit of development of various sorts on my desktop, I also wanted to muck about with containers (which I’ve done for isolating builds), but I wanted to bridge the container with my network-facing interface in a manner that it would happily pull down IPv4 and IPv6 addresses from DHCP. My network, on both IP protocols, is serviced by DHCPv4 and DHCPv6 with more-or-less static assignments tossed out to the appropriate systems with a few exceptions. Specifically, connected devices either get their automatically configured IPv6 address, a static address plus an automatically configured address, or some mix (or absence) thereof. Ideally, containers should receive these addresses as well and behave just as an ordinary OS instance or virtual machine would.

Unfortunately, if you follow the systemd-networkd guides for Arch, you’ll quickly wind up in a circumstance where your network is only going to be half-configured. You might be able to get the IPv4 addresses you expect, if you configure the bridge for DHCP, but your IPv6 range isn’t necessarily going to work even if it does accumulate a set of addresses. Bummer.

Oh, and pay attention to casing: If you enter any of the values entirely in lowercase even by accident, nothing will work. I made this mistake, and my eyes never saw anything wrong because I was too focused on reading “bridge” to notice that it should have been entered as “Bridge.” The obsession, in this case, with the word “bridge” apparently served as an override for sensible parsing that included letter case, but this is a fairly common problem when you’ve been steeped in a similar class of issues for days on multiple systems; it’s a similar phenomenon to semantic satiation wherein words begin to look “strange” if you stare at them too long.

For the purposes of this discussion, we’ll assume that we’ve created files in a similar spirit to the Arch guide for systemd-networkd bridges: vbr0.netdev, vbr0.network, and whatever the name is of your hardware interface (mine is enp6s0, so I named its configuration rather creatively: enp6s0.network).

As with the guide, our files are as follows:

# vbr0.netdev
[NetDev]
Name=vbr0
Kind=bridge
# vbr0.network
[Match]
Name=vbr0
 
[Network]
DHCP=yes

(Notice DHCP=yes above. If you need static assignments, change this accordingly.)

And finally:

# enp6s0.network
[Match]
Name=enp6s0
 
[Network]
IPv6AcceptRouterAdvertisements=no
Bridge=vbr0

Oh! What’s this IPv6AcceptRouterAdvertisements=no? More importantly why do we need it? We want to accept IPv6 RAs on our network, don’t we?

The answer might surprise you: Yes and no. Yes, we need router advertisements, but no, we don’t need them on the bridge’s slave device. (Bonus: You won’t find this in the current documentation.) If you fail to add this option, your physical ethernet device will collect an autoconfiguration address, probably accumulate the appropriate route information, configure itself for IPv6, all while your bridge device does the same thing. You’ll be left with an interface that works for IPv4, is configured for both IPv4 and IPv6, but refuses to do anything with its IPv6 assignments. Obviously, if your network isn’t running radvd or similar, you won’t need this, and you certainly won’t need this if you’re not using IPv6. However, if you’re not running IPv6, you probably aren’t reading this article, are you?

If you’ve made it this far, I assume you’re interested in the final moment of truth and the culmination of our journey together. I therefore present to you a systemd-nspawn container with DHCP assignments for IPv4 and IPv6 auto configuration (addresses redacted for privacy):

[root@buildbot network]# ip addr
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default qlen 1
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet 127.0.0.1/8 scope host lo
       valid_lft forever preferred_lft forever
    inet6 ::1/128 scope host 
       valid_lft forever preferred_lft forever
2: host0@if5: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default qlen 1000
    link/ether 1e:21:2b:fc:51:b9 brd ff:ff:ff:ff:ff:ff link-netnsid 0
    inet 192.168.5.24/23 brd 192.168.5.255 scope global dynamic host0
       valid_lft 938sec preferred_lft 938sec
    inet6 [prefix redacted]:664e:9f3d/128 scope global noprefixroute 
       valid_lft forever preferred_lft forever
    inet6 [prefix redacted]:fefc:51b9/64 scope global mngtmpaddr noprefixroute 
       valid_lft forever preferred_lft forever
    inet6 fe80::1c21:2bff:fefc:51b9/64 scope link 
       valid_lft forever preferred_lft forever

If you haven’t bothered planning your IPv6 migration, you really ought to. The inevitable is coming sooner or later. You do want to be prepared, don’t you?

No comments.
***

Updates Suck (Part 2): KDE Plasma 5.5 breaks task manager minimized application display

If you’ve recently updated to KDE Plasma 5.5, you may have noticed that minimizing your applications no longer produces a noticeable effect in the task manager. (In actuality, the status effect of minimized versus non-minimized applications is limited to a thin, gray outline around the application title–hardly something that conveys application state well enough to process visually at first glance.)

The fault of this change lies with this bug report posted almost 3 years ago. After much discussion, it was decided that 1) the “disabled” state of minimized applications was a mistake, 2) the minimized state makes it too difficult for some people to locate the gray scale application icon, and 3) no one wanted to add a configuration option to enable or disable this feature, so the next best thing was to rip it out. Now, I’ll be honest: Maybe I just happen to have better-than-average pattern matching capabilities in my visual cortex (doubtful), but the gray scale icons for minimized applications really don’t bother me. What does bother me is this change: It takes me twice as long to find minimized applications on the task manager. To say nothing of the fact that two or three people complaining incessantly for 3 years can effect such a substantial change on the rest of us who actually liked the feature…

Humorously, I didn’t notice this change during the initial update. Since I spread my work out across multiple desktops, I don’t often minimize my applications. This explains, at least in part, why it took three days for me to discover this problem and make a post about it (and why I blamed it on plasma-shell’s propensity to crash all the time, although that seemed like a reasonable explanation in lieu of anything more substantive).

The good news is that I traced it down to this commit which turned off the minimized gray scale display. As expected, removing this patch returns the task manager’s display of minimized applications to its previous state. The bad news is that, yes, you do need to recompile plasma-desktop from source. If you happen to use Arch Linux, I’ve put together a PKGBUILD containing the fix based off the PKGBUILD in ABS (hence why I haven’t yet changed any of the credits–I should note that this is not an official package). I’ll keep this updated as best as I can until such time that upstream finally settles on a better idea than a 1 pixel gray outline for non-minimized applications.

The rest of you will have to apply the patch and build the sources yourself. Sorry about that.

No comments.
***
Page 1 of 4912345...102030...Last »