Flash memory generally doesn't have great lifespan for repeated writes; using it as a root device is going to kill it pretty quickly. I had a similar problem once running a NAS off a USB stick; within a year of so the stick would no longer accept new data and I swapped it out for an SSD instead.
I suspect most of the "corruption" RPi users experience is likely bad power (shitty undersized USB power bricks) glitching the CPU and causing the filesystem to corrupt. The SD card is likely doing exactly as it's being told by the host system - write (invalid) data to blocks, which it does.
It's actually weird that Raspberry Pis seem so bad at using SD cards.
I've just scrapped a large PBX which had a whopping 8MB SD card in it. Its 486 processor booted Linux off the SD card, it wrote CDRs to the SD card, it served said CDRs up over a web interface off it, it saved its logs to the card, and all the other day-to-day chatter of a running system.
It was installed pretty much 20 years ago. The date sticker on the SD card shows that it was replaced 18 years ago, presumably as part of an upgrade.
I'm going to guess that it was not a normal COTS SD card, and instead designed for continuous writes.
The factors for SD card failures in raspberry pi seem to be mostly SD cards not meant for the purpose and having things like swap, atime or disk logs enabled which leads to unnecessary writes.
We “lost” a signage device in a facility we were doing an IoT refresh on.
It was an old wyse terminal, and at some point a CF adapter with an SD card was added, and a construction project buried it inside of a block wall. lol. It’s still there.
I’ve had the same issues in the Pi1-3 era, can’t say for sure if it was the Pi, the SD card, or the software running on top (HA was a big culprit). The last issues were probably before 2018-2018. Since then I don’t remember having any failure and I have a bunch of Pi3s and 4s running with flawlessly for about that long, including HA.
Since then I’ve been careful to buy high quality cards, oversized them a bit, optimize the OS and apps with regards to writes (like no unnecessary logging), and paid attention to airflow. The software probably also evolved to fix the issues.
I have a rasperry pi 2 and 4 that have been running on the same size uSD (reputable brand) for 5 years and haven’t had any issues. I have swap turned off. You just have to be aware of the i/o and not do anything stupid. Clearly for heavy i/o with multitasking they are not ideal.
You can buy cards that support SMART if you want to track wear, but they are expensive.
My 3 Pis have destroyed 6 SD cards. One of them also somehow drew too much power when trying to boot and got so hot it burnt me and melted itself but that could have been the power supply. I know the 4s and 5s can blow SD cards if you run them without heatsinks.
What I'm saying is, your anecdotal evidence of two pis doesn't justify saying we did something wrong with the configuration. In my experience raspberry pis still behave like "throwaway devices" but their price tags do not justify it anymore.
"Raspberry Pis are notorious for SD card corruption with unknown causes."
This is not specific to RPis. When an SD card is subjected to "unpredictable" writes it creates potential risk of corruption. The writes are triggered by software, not the RPi hardware
There is no rule that says the RPi user must mount the root filesystem r/w on the card. It can be mounted r/w on mfs or tmpfs, for example, and the card can be removed after boot. Been doing this since 2012
Zero writes to the card, no corruption
It's true these corruption issues are "notorious" but that's due to RPi owner behaviour, not the RPI hardware
And when you say "due to RPi owner behaviour" what you're saying is "due to the default setup". It's a Raspberry Pi problem. The devices are notorious for this issue for good reason.
> Raspberry Pis are notorious for SD card corruption
True SD cards are less than ideal.
But also I suspect half the problem specifically with Pi and SD cards is that people use cheap-ass SD cards and maybe ones they found in the bottom of a drawer that may or may not have previously been used in another device (e.g. camera).
I suspect if people bought industrial SD cards instead of consumer-grade junk they might get a better lifespan out of them.
I used to get Sandisk cards but recently I stopped using SD cards altogether. When it came time to update from Debian 12 to 13 for my Pi4 I migrated to an NVME USB enclosure for my boot drive. Just yesterday I set up another Pi4 for a project with an M.2 SATA enclosure and in both cases they work just fine.
I discovered that barely-used enterprise Intel SSDs (DC S3610, S3700, S3710) were cheap on ebay a year or so ago and I may have bought a few dozen. These things have remaining endurance measured in petabytes. Sure, it’s overkill to have a 200GB enterprise SATA disk as the boot volume for a Raspberry Pi, but this is exactly my kind of overkill.
Could be some other hardware issue for you, yeah. I recall there are some problems that can happen if the usb power supply is too weak. But this fixed it for me.
Yeah that's more or less my goal for the setup in the post... I didn't have much time to investigate what else I could move off of the SD card but I suspect there's a few other things I could do. NixOS makes it super easy too.
Can confirm, I've got a pi 5 here that's been running 24/7 for over a year (possibly two)... it boots fine, but anything else makes it run so slowly because the sdcard is trashed.
The current setup writes to the systemd journal at least twice every minute... It's next iteration is getting a proper nvme.
Ho-ho my friend! You want to buy storage for your computer before 2030? I'm so sorry, but you now pay big time, now you sell your car for one nvme drive.
Really would just prefer a working, non-buggy, Pi2 maybe without all the crazy extra power + compute + RAM. Or at least have the option, because $20 is more my style for a lot of these tasks, not $150CDN.
Classic power failure and raspberry pi with SD cards. When I last set up things for my parents, I simply imaged a bunch of SD cards and had the data stored on an external hard drive. I live in the US and they live in India and this is for their RaspBMC. With the SD cards they just swap a broken one with a new one and call it a day.
Been 12 years or so now. Went back to their home a few years ago and everything but the opensubtitles integration works. I was surprised. I could never recreate it now. Even has one of these flirc setups so the remote works.
India has unreliable power too so the SD card death was inevitable. I made a few more images and called it a day.
My Raspberry Pi 4 has been used to run both a Jitsi and Trackmania server over the years. I never considered using the SD card since it can boot and run from a USB 3 flash drive. Later on, I picked up an Argon One case to add a SATA SSD.
Thankfully, modern Rockchip SBCs tend to come with NVMe slots.
The main point is to use zswap over swap plus zram. The sd card usecase definitely doesnt want swap to disk, compressed or otherwise, so zram or no swap are the options.
Moving swap off of sdcard is indeed a good plan.
But since swapping is done when ram is depleted, using a ramdisk for it (which uses ram) is ... exactly what you dont want.
Since OP states they have disks attached to the machine as well, the better thing is use a partition there for swap.
I use a Waveshare hat for my Pi 5s. It does PoE and has a spot for an SSD. What I particularly like about this setup is the single wire running to each board, looks very clean.
I'm not a sysadmin but I often hear about server failures due to hardware just wearing out. Is there an expected shelf life of the hardware? I would expect for sure on the storage devices. So if that were the case then why isn't there some sore of daemon or system software service that tracks the expected life of the hardware? I would expect that the software would start showing warnings of imminent useful life of devices to expire.
My primary home zfs server is from 2010, running 24x7 since then. It has a 4-way mirror zfs pool for data. Over the years I have replaced two of the four drives, so two are still original from 2010. Everything else on the machine is from 2010.
You can get unlucky and have components fail quickly, but usually if they last, they'll last for a long time.
But trying to run a server from an SD card like in the article, seems silly. False economy, just get a good SSD.
For drive health monitoring there is SMART, but IME while it does monitor drive usage and lifetime, it rarely predicts catastrophic failure since that happens suddenly. Both drives that failed in this system went from healthy (according to SMART) to unresponsive overnight.
There is for hard drives - SMART - the problem is, unless you're an enterprise (and let's all stop and laugh because we KNOW that most "enterprises" run at least some hardware until the only person who even knows it exists is long in the grave) you're not going to do preventative replacement.
And especially for home use, where nien 9s is more realistic than nine 9s, you're better off having a solid backup strategy and a source of spare parts.
My main "tank" server is 16 years old this year, and has blown out one motherboard - it was a bit annoying to be "offline" for a few days as a replacement wandered in from eBay, but it worked. If it happened today I might import the zpool into another device temporarily - or permanently, who knows? This R510 probably takes more power than a datacenter.
I run a k8s cluster in my homelab, with Longhorn providing storage. Except for some extra-large volumes for bulk storage (just movies in Jellyfin right now), everything is replicated three ways. I've had nodes fail to boot up after a power outage (turned out to be a dead CMOS battery) and procrastinated bringing them back up because everything just kept working.
I run a fairly large homelab (probably ~10 nodes when including OPNSense routers, backup NAS at our cottage, etc.), and hardware failures have been incredibly rare, especially compared to what it used to be a couple of decades ago (I had quite a few motherboards fail due to blown capacitors due to the stolen electrolyte formula scandal).
Hardware failures tend to follow a bathtub curve: there are more frequent failures early on due to manufacturing defects, then they become rare until you reach the end of the natural lifespan of the equipment. That is shorter for some components (spinning rust HDDs, fans, CMOS batteries, power supplies in noisy environments), but extremely long for others.
I'm running my own home server on an old 2014 Mac mini, 24/7 basically. It both has a SSD and a HDD (old "Fusion" disk type of machine, but I use both parts separately, the SSD part for the OS, the HDD part for the data.)
A few days ago the HDD had its 100,000 Power_On_Hours anniversary. :) According to SMART this is way past the expected/ designed life time of the disk (the normalised value reports "1" since a few years when the disk hit about 42,000 Power_On_Hours)
So you never know. It either works or it doesn't. You have backups for the worst case.
My main desktop machine with assorted components is from 2014, also with an SSD and HDD. Each component has lasted all this time somehow. The total uptime of each drive is 11.5 years (101,023 hours) with 159 power cycles. The SSD is 256 GB and has 71 TB written, it's my main drive. It has 46% lifetime remaining, I hope everything lasts another 10+ years.
> So if that were the case then why isn't there some sore of daemon or system software service that tracks the expected life of the hardware?
Some linux distribution will regularly check SMART data from disk and warn you if some threshold has passed (like number of corrected errors, spin up/down, total hours etc).
That said, if we talk about consumer hardware in a home lab or the like, of the disks that I bought in the past 15 years, only one broke down, and it did fairly quickly so my suggestion is to use redundant storage, make backup and change the disk when it breaks down. (Enterprise disk and usage is another can of worms of course).
Life expectancy really depends on how the hardware gets used. I run a small fleet of servers that are heavily CPU / GPU loaded, and the most frequent hardware failure I see is power supplies, with RAM as #2 and spinning HDDs as #3. Each of these (on enterprise-grade gear) has some amount of reporting / monitoring, but it's rarely the case that I get any advance warning before a failure. The RAM is probably the nicest, as it sometimes starts with correctable ECC warnings. Nearly all of the PSU failures have been sudden deaths, with the only saving grace being N+1 redundant supplies usually keeping the victim server alive.
About the only thing that regularly wears out is spinning drives, and those have SMART data to give you an idea of when they'll die, but generally yif you're actually committing to high uptime you'll have redundant hardware so any one server failing doesn't take the service down, giving a bit of a grace period on hardware failures
Hardware wearing out is not common at the individual component level. It only becomes common when you are running a large number of servers, those individual small probabilities add up.
> Not eager to ruin the evening by starting a debugging session
Use a clanker. Nothing has brought back so much joy into homelabbing for me as having LLMs handle the boring stuff.
I launch it in opencode on the system or elsewhere to ssh in (I think giving it KVM is an overkill but an option nonetheless) and tell it to fix things. Not just ask questions and generate configs. I have backups, let it rip. They have gotten shockingly good at it even when they write bash wrappers just to catch some logs.
If i'm reading this correctly, can run Immich from a Raspberry Pi 4? That's pretty cool.
I bought all the stuff (new case, networking card and a Jonsbro N4 to repurpose some intel i7-9700 era processor and motherboard, but didn't get it setup yet, as I was starting to rethink the power consumption....
Also looking to use this rebuild as an opportunity to learn Nix.
Let me know if you all have some good article references, to point me at. I've been doing some research, but could always do more.
I may end up putting a low power NUC in the case, or something. I don't know...
You can, I am self-hosting Immich via Docker on Raspberry Pi 4 with small SSD and it's nothing but good experience so far (but I keep the photos backed up elsewhere just in case, I use Immich mostly to share curated sets of photos)
> If i'm reading this correctly, can run Immich from a Raspberry Pi 4? That's pretty cool.
Yes! It runs completely fine on a Pi, except for the machine learning service. The good news is that you can run that service on a separate computer with a GPU and point Immich to it. Immich will use it when it's available and I think you can configure it to run all the queued ML jobs at a certain time when you expect the computer running the ML service to be on.
Lack of proper IO connector is the reason I abandoned Raspi. SD card crashes are nearly a guarantee. Total pita too since there’s no display during the BIOS.
Another useful advise would be to keep your boot drive 60% used at max and run TRIM after freeing a huge chunk of files to let the wear leveling algo do its job.
And every few years reinstall the OS to free up unwritten cells.
Personally I would get a used Dell Optiplex / Lenovo ThinkCentre / HP EliteDesk. Either the "mini" / "micro" if you really want small. I am a fan of the SFF (Small Form Factor) size and the Dell Optiplex models. In particular the 70XX versions that offer Intel AMT / vPro for remote management.
I run a Proxmox server on my Dell Optiplex SFF 7070. With the SFF size you can add two low-profile PCIe cards, if desired. In my case I added a 10Gbit NIC.
The reason I would go with a used Dell / Lenovo / HP is that these tend to be well-built machines that run for years and years 24/7 in offices and proven to work well.
I picked Beelink a couple years ago. It didn’t last long. I can’t simply reboot it. I have to unplug it, wait 5 minutes, plug it in, and then boot it. It turns OS installs and even system updates into such a hassle that it’s not even worth using anymore. I wanted to run it headless and just connect remotely, and that doesn’t really work well with this issue, as I need regular physical access.
I assume it’s a bad capacitor or something. I tried changing out the power adapter, but that didn’t help.
It really is a high failure rate.
I've just scrapped a large PBX which had a whopping 8MB SD card in it. Its 486 processor booted Linux off the SD card, it wrote CDRs to the SD card, it served said CDRs up over a web interface off it, it saved its logs to the card, and all the other day-to-day chatter of a running system.
It was installed pretty much 20 years ago. The date sticker on the SD card shows that it was replaced 18 years ago, presumably as part of an upgrade.
The factors for SD card failures in raspberry pi seem to be mostly SD cards not meant for the purpose and having things like swap, atime or disk logs enabled which leads to unnecessary writes.
It was an old wyse terminal, and at some point a CF adapter with an SD card was added, and a construction project buried it inside of a block wall. lol. It’s still there.
Let alone things like putting /var/tmp and/or /var/log in RAM, or long commit intervals.
No idea what the current situation is; Pis are too dear for the likes of me now.
relatime is a default for a while now, like decade+, you don't need to set up mount option just to get relatime
Since then I’ve been careful to buy high quality cards, oversized them a bit, optimize the OS and apps with regards to writes (like no unnecessary logging), and paid attention to airflow. The software probably also evolved to fix the issues.
You can buy cards that support SMART if you want to track wear, but they are expensive.
What I'm saying is, your anecdotal evidence of two pis doesn't justify saying we did something wrong with the configuration. In my experience raspberry pis still behave like "throwaway devices" but their price tags do not justify it anymore.
This is not specific to RPis. When an SD card is subjected to "unpredictable" writes it creates potential risk of corruption. The writes are triggered by software, not the RPi hardware
There is no rule that says the RPi user must mount the root filesystem r/w on the card. It can be mounted r/w on mfs or tmpfs, for example, and the card can be removed after boot. Been doing this since 2012
Zero writes to the card, no corruption
It's true these corruption issues are "notorious" but that's due to RPi owner behaviour, not the RPI hardware
True SD cards are less than ideal.
But also I suspect half the problem specifically with Pi and SD cards is that people use cheap-ass SD cards and maybe ones they found in the bottom of a drawer that may or may not have previously been used in another device (e.g. camera).
I suspect if people bought industrial SD cards instead of consumer-grade junk they might get a better lifespan out of them.
Makes it easy to replace, though.
The current setup writes to the systemd journal at least twice every minute... It's next iteration is getting a proper nvme.
Somebody set up us the bomb.
Rpi5 has a pcie slot if you want a cleaner setup.
Been 12 years or so now. Went back to their home a few years ago and everything but the opensubtitles integration works. I was surprised. I could never recreate it now. Even has one of these flirc setups so the remote works.
India has unreliable power too so the SD card death was inevitable. I made a few more images and called it a day.
Thankfully, modern Rockchip SBCs tend to come with NVMe slots.
every other component is cheap enough but not Ram.
swap is used when you run out of ram. using a ramdisk for swap ... You see how that doesnt make sense?
You can get unlucky and have components fail quickly, but usually if they last, they'll last for a long time.
But trying to run a server from an SD card like in the article, seems silly. False economy, just get a good SSD.
For drive health monitoring there is SMART, but IME while it does monitor drive usage and lifetime, it rarely predicts catastrophic failure since that happens suddenly. Both drives that failed in this system went from healthy (according to SMART) to unresponsive overnight.
And especially for home use, where nien 9s is more realistic than nine 9s, you're better off having a solid backup strategy and a source of spare parts.
My main "tank" server is 16 years old this year, and has blown out one motherboard - it was a bit annoying to be "offline" for a few days as a replacement wandered in from eBay, but it worked. If it happened today I might import the zpool into another device temporarily - or permanently, who knows? This R510 probably takes more power than a datacenter.
Hardware failures tend to follow a bathtub curve: there are more frequent failures early on due to manufacturing defects, then they become rare until you reach the end of the natural lifespan of the equipment. That is shorter for some components (spinning rust HDDs, fans, CMOS batteries, power supplies in noisy environments), but extremely long for others.
A few days ago the HDD had its 100,000 Power_On_Hours anniversary. :) According to SMART this is way past the expected/ designed life time of the disk (the normalised value reports "1" since a few years when the disk hit about 42,000 Power_On_Hours)
So you never know. It either works or it doesn't. You have backups for the worst case.
Some linux distribution will regularly check SMART data from disk and warn you if some threshold has passed (like number of corrected errors, spin up/down, total hours etc).
That said, if we talk about consumer hardware in a home lab or the like, of the disks that I bought in the past 15 years, only one broke down, and it did fairly quickly so my suggestion is to use redundant storage, make backup and change the disk when it breaks down. (Enterprise disk and usage is another can of worms of course).
Back in the day I ran s××× coin miners off flash drives without any apparent problems.
Teslas though... ;)
Use a clanker. Nothing has brought back so much joy into homelabbing for me as having LLMs handle the boring stuff.
I launch it in opencode on the system or elsewhere to ssh in (I think giving it KVM is an overkill but an option nonetheless) and tell it to fix things. Not just ask questions and generate configs. I have backups, let it rip. They have gotten shockingly good at it even when they write bash wrappers just to catch some logs.
I bought all the stuff (new case, networking card and a Jonsbro N4 to repurpose some intel i7-9700 era processor and motherboard, but didn't get it setup yet, as I was starting to rethink the power consumption....
Also looking to use this rebuild as an opportunity to learn Nix.
Let me know if you all have some good article references, to point me at. I've been doing some research, but could always do more.
I may end up putting a low power NUC in the case, or something. I don't know...
Yes! It runs completely fine on a Pi, except for the machine learning service. The good news is that you can run that service on a separate computer with a GPU and point Immich to it. Immich will use it when it's available and I think you can configure it to run all the queued ML jobs at a certain time when you expect the computer running the ML service to be on.
I now use a small SSD connected via USB3 and a slightly modified version of this : https://framps.github.io/raspiBackupDoc/introduction.html
...to take snapshots to my NAS periodically.
I keep a similar small SSD in the drawer ready for a restore and swap if the main system drive fails.
And every few years reinstall the OS to free up unwritten cells.
I run a Proxmox server on my Dell Optiplex SFF 7070. With the SFF size you can add two low-profile PCIe cards, if desired. In my case I added a 10Gbit NIC.
The reason I would go with a used Dell / Lenovo / HP is that these tend to be well-built machines that run for years and years 24/7 in offices and proven to work well.
I assume it’s a bad capacitor or something. I tried changing out the power adapter, but that didn’t help.
I don’t see myself ever buying another Beelink.