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).


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:


  • 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)


  • 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.

To 404 or not to 404?

If you’ve been following my (admittedly) rare posts over the course of the last few years, you’re likely to have noticed a growing aggression toward PHP as a language. It’s not that I hate the language per se (although I sometimes do), it’s that there’s so much crap written in PHP that it’s almost impossible to find something well-written that’s pleasant to work with (major props to Fabien Potencier of SensioLabs for writing some of the best PHP I’ve ever had the pleasure to work with–see, I can be even-handed!). There are some positive developments, particularly in commercial PHP, but all things considered, we’ve a long way to go.

XenForo (as an example) is leaps and bounds better than vBulletin in terms of actual design. They use a well-tested framework (Zend), mostly keep to the philosophy of separation of concerns, and isolate components from each other without polluting the global namespace (hi, vBulletin!). However, (you were waiting for this, weren’t you?) the coding style still remains abysmal (what’s wrong the K&R style, guys? seriously!) and it’s almost infuriating how much magic is used to create classes on the fly for various tasks. Need an example? Just look through some of the deferred tasks: Nearly everything is generated by factory classes that accept a class to instantiate as a string argument, then return the instantiated class. This absolutely screws with a developer’s ability to rationalize about types being passed around, and if it weren’t so infuriating, I think most of us would break down in tears. Do you want to see a grown man cry?

Yes, I can understand why the XenForo developers have made their choices (no need for a long switch statement, object selection for instantiation becomes the concern of the calling code, etc.), but it’s absolutely a tremendous pain in the arse when one happens to be trawling through the sources to figure out why something isn’t working. Let’s not even get into the whole bazillion singleton methods splattered everywhere like the ground zero of a misplaced spittoon. How’s that for a visual? I think I just gave myself a fist-bump.

But alas, I distract myself with unnecessary things. This post is supposed to be about 404 errors, isn’t it? And here I was about to launch into an argument about how there’s so many libraries out there that support dependency injection, you’d have to be insane to use a hundred different singleton classes.

Today I encountered a relatively curious thing with XenForo. John (of Forum Foundry, Inc.) was having some difficulty with the nginx configuration on one of his sites and had convinced himself it was his fault because of something he had changed (it wasn’t). Something, somewhere was causing a 404 page to crop up every time users would attempt to reset their passwords. It was bugging him, and he asked if I’d take a look. So I did.

After examining his nginx configuration, I was convinced he did nothing wrong. But disconcertingly, I grew increasingly more convinced that it wasn’t nginx’s fault, either. Strangely, though, the 404 page was being spat out by nginx–not by XenForo–so it seemed impossible for the fault to lay elsewhere. It had to be nginx. Yet, I found myself battling with a sort of cognitive dissonance: Posting (using curl or similar) to the lost-password URL returned a 200 OK. Using GET (or other methods) returned a 405 Method Not Allowed (as expected). If it were the fault of nginx, the 404 error should be returned regardless of the method (GET, POST, or otherwise).

We tried various things, mostly involving the nginx debug log, but couldn’t quite get a handle on the source of the problem. I was pretty sure PHP was somehow at fault, but I couldn’t duplicate bubbling a 404 error up from PHP to nginx. Puzzled, I sat back and thought about the problem, and on a whim decided to do some digging around. It wasn’t long before I stumbled on the documentation for nginx’s fastcgi_intercept_errors directive, which seemed as if it might exacerbate upward error propagation, from application code. From the documentation:

[fastcgi_intercept_errors] [d]etermines whether FastCGI server responses with codes greater than or equal to 300 should be passed to a client or be redirected to nginx for processing with the error_page directive.

What does this mean? It means that, if fastcgi_intercept_errors is enabled, HTTP response codes with a value of 300 or greater (of which 404 is one) will trickle up and be intercepted by nginx regardless of the application (PHP) layer’s intent. Thus, even if the application displays something useful with the 404 message, nginx will intercept the response and display its own spartan 404 page.

Then a lightbulb came on.

Around line 63 or so of the XenForo lost-password handler, a 404 response is generated if a user’s name or email address doesn’t exist when a request is submitted to reset their password. This means that, in all likelihood, if someone mistypes their username (or email), they’ll receive a 404 message from nginx and no error message from XenForo. Rather than attempting to hit the back button, change their submission, and try again, the affected user is likely to believe the site is broken and they are therefore unable to reset their password. This couldn’t be further from the truth.

But it gets better. I didn’t stumble on fastcgi_intercept_errors initially, because I couldn’t get the 404 responses to bubble up to nginx for interception on my system (Arch Linux). The reason for this is that the /etc/nginx/fastcgi_params that ships with Arch is fairly sparse (the intent being that if you need more specific parameters, you add them yourself). But on this server, that wasn’t the case, because fastcgi_intercept_errors was enabled. Moreover, tagging along with it was a number of other parameters (I initially though it was the fault of Ubuntu), but it became clear that these were not the defaults that shipped with nginx (it doesn’t ship with fastcgi_intercept_errors enabled, for instance) or Ubuntu. Thus, any time a 404 error was dispatched from the application code, nginx intercepted it and displayed its 404 error page instead.

How many others might be affected by this? Who knows. If there’s a tutorial out there that recommends adding fastcgi_intercept_errors to /etc/nginx/fastcgi_params (it turns out there might be–read the update below), it goes without saying that the suggestion is a terrible idea if you expect your applications to generate and handle 404 errors (or greater). Of course, for high performance configurations, you probably want nginx to do as much of the heavy lifting as possible, but in this particular circumstance, we encountered an issue where using certain configuration options can yield unexpected results. In other words, don’t blindly copy and paste “suggested performance tweaks” from random Internet blogs without first consulting the documentation. The results may surprise you. (I don’t know if that’s what happened in this case, but I certainly can’t rule it out. For all I know, it may have been–and probably was–an accident.)

Is this the fault of XenForo? I’d argue no. While sending a 404 error in response to members that don’t exist who have attempted to reset their passwords is probably intended to be a RESTful response (not a member? respond with a 404) that does something unexpected (nginx 404 instead), it’s not a bug per se. However, it’s a potential privacy concern: Anyone who has access to a large corpus of email addresses can deduce based on the response returned by XenForo whether or not those email addresses are registered with the server. By doing so, a potential attacker can then target those accounts that are known to exist on the server for nefarious deeds (or violate their privacy).

But here is where we diverge into a matter of directed concerns: Do you provide users (some of whom might be terrible typists) with immediate feedback, indicating that the information they submitted was wrong, thereby increasing the usability of the site; or do you protect users’ privacy by generating the same response whether or not the information was submitted correctly? There is no right or wrong answer to this question, and generally it boils down purely to the matter of usability. Do you value ease of use or privacy? You can’t have both. Therefore, the answer rests on the shoulders of the site operator to decide. For sites that value users’ privacy, perhaps it’s better to sacrifice some usability for security.

For everyone else, it’s probably best to leave /etc/nginx/fastcgi_params at its default unless you absolutely must change the defaults. Better: Place the tweaks you need in a separate file and include that file separately.

Update I think I may have discovered one of the culprits for fastcgi_intercept_errors here, listed as the “Ultimate Speed Guide for WordPress on NGINX.” Hint: Don’t blindly copy everything someone on the Internet tells you is a good idea. If you don’t know what something does, always read the documentation. If you don’t, it will bite you.

No comments.

Firefox Fonts too Light (or blurry)? Here’s a Fix

It’s no secret (or surprise) that Mozilla released Firefox 6 today. Included in this release is, among other things, working WebGL support and numerous fixes. There’s one small, glaring annoyance that some of you on Windows might have noticed:

While that snapshot doesn’t do much justice to the degree that Firefox 6.0’s fonts are most certainly light, other parts of the UI suffer. I’m not precisely sure what the culprit is; it vaguely reminds me of the rather light default font rendering that shipped with WPF in Visual Studio 2008.

Anyway, one possible fix (I’m sure there’s more) dates back some 5 months prior and can be had by entering about:config into a new tab, filtering for gfx and changing the values gfx.direct2d.disabled and gfx.direct2d-force-enabled to true. This might affect performance and possibly introduce breakage, but it works fine under Windows 7. If you’re curious, here’s what the appropriate settings should look like when you’re done:

This solution will probably not work on older versions of Windows that did not use a GPU-accelerated window manager, and I’m not sure how it will behave in Windows 7 with Aero turned off. Be mindful of this. You might break things.

Hope that helps!