answered 2018-01-22 13:11:50 +0200

DrYak
9624 ●93 ●134 ●128

http://www.sympato.ch/

First big caveat :

• in my tests, the MMC port used by the SDCards is slower than the eMMC port used by the build-in memory

  (Meaning that even if you go for a UHS Speed Class 3 - the U3 logo on cards - you won't be able to achieve the maximal 90 MiB/s bidirectionnal advertised by the manufacturer.

that said...

Card Reliability :

• aim for a card that provide static wear levelling in addition to the standard dynamic wear levelling.

  i.e.: Cards that also automatically wear-level and relocate data that is already stored on the card. On a 128GiB card, you're going to have quite a lot of data, but not all of it is going to be regularily over-written (dynamic). Some data stays the same untouched (static) and if left on the cells, there's a tiny risk of it going stale and bitrot (corruption happening) and/or the distorted cell levels giving more read error and the read speed suddenly dropping to a crawl (that's a problem that was even observed on Samsung 840 EVO SSD with early generation firmware)

• aim for a card that features ECC (i.e.: error correction) so even in case of data corruption, it might manage to recover your data.
• When possible, go for industrial variant of the card (they are usually certified for higher reliability) but it might not be available in the sizes that interests you (often, reliability is provided by using SLC or by using lower density MLC cells. Thus no 128GB for you).
• Check forum for how the card behaves when dying. Most of the high quality brands will lock the card into read-only mode. That means you could always attempt to recover and extract files by plugging it into a laptop with a recent enough version of BTRFS and doing btrfs restore (or sometime just plain mounting with the recovery options and using older CoW copies).

In my personal experience, I'm quite happy with Transcent's Ultimate 633x line of cards (they also have 128GiB in slightly less fast Ultimate variants), it fulfills most of my criteria (ECC, static wear leveling). I've even successfully managed to use them for swap (and quite intensively) without burning them.

Btrfs reliability :

• use RAID1 whenever possible (for metadata). If you've left the on-board BTRFS filesystem in default "dup" configuration, it will automatically upgrade to RAID-1 when you add the second volume. If you've converted to "single" (because on wear-leveled flash media, "dup" doesn't actually provide any advantage), you'll need to -mconvert.
• btrfs scrub the filesystem often (like at least weekly), that will help detecting bitrot (and depending on the firmware running on the embed chip in charge of wear levelling, it might trigger ECC corrections and rewrite).
• btrfs balance the filesystem also (e.g.: weekly). No need to go all the way to 100% like the automatic balancer that is run before system upgrades, if you go up to -musage=60 -dusage=60 that should be enough. That would help running out of free blocks (enospc) on the internal flash for the RAID-1ed metadata.
• Do. Backups. Saddly the Jolla 1's 3.2 kernel's btrfs send isn't functionnal, but you can at least setup a pedioic rsync, in case of MMC media failure.
• Do not try to fsck, or only as a last resort (after trying to recover/extract the files first with btrfs restore). Per-se, fsck doesn't make much sens on CoW (BTRFS, ZFS) and Log-structured (UDF, F2FS, etc.) filesystems. And the script provided on the Jolla Recovery image is very bad in my opinion (it tries stupid stuff like erasing the checksums).