pcouy

Thanks for sharing your experience and for the links.

Do you think it would be doable to make/host a tileserver that only generates the first few zoom levels for the whole planet by default, and is able to generate tiles for more detailed zoom levels only for specific locations ? I'm thinking of a feature where Immich asks the tile server to generate the appropriate tiles based on the locations of photos. Since we only ever zoom on locations where photos have been taken, and we often take several photos at the same locations, could this decrease the requirements enough for self-hosting ?

No need to be rude...

Thanks for the detailed feedback. According to one Immich dev, they used to use OSM's raster tile provider but switched away from it since they were causing too much load on OSM's servers.

There does not seem to be any non-commercial vector-tile provider at the moment (though OSM seems to be currently working on it), and it seems really overkill to try and self-host a tile provider (at least with the default level of details). Maybe the way is to find a balanced level of details that makes it reasonable to self host

Quoting one dev from the conversation I had on Discord :

the one run by OSM is not intended for general purpose use because that results in way too much load on their system. We used to use theirs, but as Immich grew we decided that we should relieve them of that

I guess you (and they) are talking about raster tiles, since OSM does not seem to provide vector tiles

When I mentionned that "I can confirm it is not realistic to self-host a tile provider", it's because I tried to run maptiler : it maxed out my CPU for 2 hours before my disk got filled while trying to generate the tiles from OSM data (and it was just for France)

Edit : Anyway, I don't think this should be in Immich's scope. Simply providing an easy option to switch tile providers would allow people motivated enough to host maptiler to use it

Edit bis : More details on how hard it is to host your own tile provider are available on the official OSM wiki

Sur le sujet de ce qu'est capable de faire un LLM "juste en donnant des prédictions pour le mot suivant", je te recommande une autre vidéo récente de Mr Phi sur le fait de faire jouer ChatGPT aux échecs. C'est assez impressionnant, et un poil flippant d'après moi.

Si tu veux creuser le sujet (et notamment sur ce que dit Yann Le Cun, je te recommande cette vidéo si tu comprends l'anglais

J'ai pas réussi à tout lire, la condescendance du truc m'est insupportable.

J'ai compris qu'il fallait rien esperer en voyant une citation de Raphaël Enthoven, j'ai abandonné 2 lignes plus bas

Things have been going well for me, using docker-mailserver.

I followed the setup guide, did everything in the DKIM, DMARC and SPF documentation page. The initial setup required more involvement from me than your standard docker-compose self-hosting deployment, but I got no issues at all (for now, fingers crossed) after the initial setup : I never missed any inbound e-mails, and my outbound e-mails have not been rejected by any spam filter yet.

However, I agree with everyone else that you should not self-host an important contact address without proper redundancy/recovery mechanism in case anything goes wrong.

You should also understand that self-hosting an email address means you should never let your domain expire to prevent someone from receiving emails sent to you by registering your expired domain. This means you should probably not use a self-hosted e-mail to register any account on services that may outlive your self-hosted setup because e-mail is frequently used to send password reset links.

I will definetly look into this. I've been using tube archivist for a while now, but it eats so much RAM (especially the Elastic search dependency IIRC)!

I did not try it out yet, but I will make sure I do. I love a lot of things about the approach you described

I had a great time using Qubes. It made me learn about the Xen hypervisor and CoW filesystems.

However, if OP complains about build times being too long on their CPU, I'm not sure they will get Qubes running smoothly on the same hardware. I'm especially worried about every VM besides dom0 being software rendered.

Github is not really independent from Git, it's a git provider. (you could see it as Github being to Git what Gmail is to e-mails)