IPFS article (Quentin's part)
All checks were successful
continuous-integration/drone/push Build is passing
continuous-integration/drone/pr Build is passing

This commit is contained in:
Quentin 2022-06-10 17:25:53 +02:00
parent 1bbe3f9c58
commit 5895e434e4
Signed by: quentin
GPG key ID: E9602264D639FF68
6 changed files with 131 additions and 16 deletions

Binary file not shown.

After

Width:  |  Height:  |  Size: 42 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 46 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 51 KiB

View file

@ -17,29 +17,42 @@ Still, you might want to share and collaborate with the rest of the world, and i
or by connecting it to an application that will act as a "proxy" between Garage and the rest of the world.
We refer as proxy software that know how to speak federated protocols (eg. Activity Pub, Solid, RemoteStorage, etc.) or distributed/p2p protocols (eg. Bittorrent, IPFS, etc.).-->
## Preliminary
## Some context
People often struggle to see the difference between IPFS and Garage, so let's start by making clear that these projects are complementary and not interchangeable.
IPFS is a content-addressable network built in a peer-to-peer fashion.
With simple words, it means that you query the content you want with its identifier without having to know *where* it is hosted on the network, and especially on which machine.
As a side effect, you can share content over the Internet without any configuration (no firewall, NAT, fixed IP, DNS, etc.).
It has some nice benefits: if some content becomes very popular, all people that already accessed it can help serving it, and even if the original content provider goes offline, the content remains
availale in the network as long as one machine still have it.
However, IPFS does not enforce any property on the durability and availablity of your data: the collaboration mentioned earlier is
Personnaly, I see IPFS as the intersection between Bittorrent and a filesystem. Bittorrent remains probably one of the most efficient way to deliver
a copy of a file or a folder to a very large number of people. But it lacks some interactivity: once a torrent file has been generated, you can't simply
add a file to it. By adopting a filesystem abstraction, IPFS handles such use case smoothly.
<!--IPFS is a content-addressable network built in a peer-to-peer fashion.
With simple words, it means that you query the content you want with its identifier without having to know *where* it is hosted on the network, and especially on which machine.
As a side effect, you can share content over the Internet without any configuration (no firewall, NAT, fixed IP, DNS, etc.).-->
<!--However, IPFS does not enforce any property on the durability and availablity of your data: the collaboration mentioned earlier is
done only on a spontaneous approach. So at first, if you want to be sure that your content remains alive, you must keep it on your node.
And if nobody makes a copy of your content, you will loose it as soon as your node goes offline and/or crashes.
Furthermore, if you need multiple nodes to store your content, IPFS is not able to automatically place content on your nodes,
enforce a given replication amount, check the integrity of your content, and so on.
*Note: the IPFS project has another project named [IPFS Cluster](https://cluster.ipfs.io/) that addresses these problems.
We have not reviewed it yet but its 2 main differences with Garage are that 1) it does not expose an S3 API and 2) has a different consisenty model.
In the following, we suppose that these differences prevent you from deploying it in your infrastructure.*
enforce a given replication amount, check the integrity of your content, and so on.-->
However, you would probably not rely on Bittorrent to durably store your encrypted holiday pictures you shared with your friends, and this is the same for IPFS.
If at some time, everyone has a copy of the picture on their hard disk, people might delete it after a while without you knowing it.
You also can't easily collaborate to share this common treasure, for example, there is no automatic way to say that Alice and Bob
are in charge of storing the first half of the archive while Charlie and Eve are in charge of the second half.
➡️ **IPFS is designed to deliver content.**
Garage, on the contrary, is designed to spread automatically your content over all your available nodes to optimize your storage space. At the same time, it ensures that your content is always replicated exactly 3 times across the cluster (or less if you change a configuration parameter!) on different geographical zones (if possible). To access this content, you must have an API key, and have a correctly configured machine available over the network (including DNS/IP address/etc.). If the amount of traffic you receive is way larger than what your cluster can handle, your cluster will become simply unresponsive. Sharing content across people that do not trust each other, ie. who operate independant clusters, is not a feature of Garage: you have to rely on external software.
*Note: the IPFS project has another project named [IPFS Cluster](https://cluster.ipfs.io/) that allow servers to collaborate on hosting IPFS content.
The Bittorrent project created Bittorrent Sync, renamed later [Resilio](https://www.resilio.com/individuals/), that provide a filesystem abstraction based on the Bittorrent protocol.
Reviewing these solutions is out of the scope of this article, feel free to try them by yourself!*
Garage, on the contrary, is designed to spread automatically your content over all your available nodes to optimize your storage space. At the same time, it ensures that your content is always replicated exactly 3 times across the cluster (or less if you change a configuration parameter!) on different geographical zones (if possible).
<!--To access this content, you must have an API key, and have a correctly configured machine available over the network (including DNS/IP address/etc.). If the amount of traffic you receive is way larger than what your cluster can handle, your cluster will become simply unresponsive. Sharing content across people that do not trust each other, ie. who operate independant clusters, is not a feature of Garage: you have to rely on external software.-->
But independant Garage nodes can not collaborate to deliver popular contents: every one is playing alone.
All ressources created (keys, files, buckets) are tightly coupled to a cluster, people from different clusters can't collaborate on the same data (without additional software).
➡️ **Garage is designed to durably store content.**
@ -80,17 +93,24 @@ At the same time, I was monitoring Garage (through [the OpenTelemetry stack we h
Just after launching the daemon and before doing anything, we have this surprisingly active Grafana plot:
![Grafana API request rate when IPFS is idle](./idle.png)
*Legend: y axis = requests per 10 seconds, x axis = time*
It means that in average, we have around 250 requests per second, mainly to check that an IPFS block does not exist locally.
But when I start interacting with IPFS, values become so high that our default OpenTelemetry configuration can not cope with them.
These requests are triggered by the DHT service of IPFS: my node being reachable over the Internet, it acts as a public DHT server and start answering global
block requests over the whole network. Each time it sees a block request, it sends a request to our backend to see if it exists.
*We will try to tweak the IPFS configuration later - we know that we can deactivate the DHT server. For now, we will continue with the default parameters.*
When I start interacting with IPFS by sending a file or browsing the default proposed catalogs (ie. the full XKCD archive), values become so high that our default OpenTelemetry configuration can not cope with them.
We have the following error in Garage's logs:
```
OpenTelemetry trace error occurred. cannot send span to the batch span processor because the channel is full
```
It is useless to change our parameters to see what is exactly the number of requests done on the cluster: it is way too high, multiple orders of magnitude too high.
As a comparison, this whole webpage, with its pictures, triggers around 10 requests on Garage to load, and I expect seeing the same order of magnitude with IPFS.
It is useless to change Garage parameters to see what is exactly the number of requests done on the cluster: it is way too high considering my targeted usage: sharing a picture.
As a comparison, this whole webpage, with its pictures, triggers around 10 requests on Garage to load, not thousands.
I think we can conclude that this first try was a failure.
The S3 datastore on IPFS does too many request and would need some important work to optimize it.
@ -98,5 +118,100 @@ But we should not give up too fast, because Peergos folks are known to run their
## Try #2: Peergos over Garage
[Peergos](https://peergos.org/) is designed as an end-to-end encrypted and federated alternative to Nextcloud.
Internally, it is built upon IPFS and is known to have an [integration of the S3 API](https://peergos.org/posts/direct-s3).
One important point of this integration is that your browser is able to bypass both the Peergos daemon and the IPFS daemon
to write and read IPFS blocks directly on the backemd.
<!--So let's see if we are able to connect Peergos on Garage and have some federation across the Internet through IPFS.-->
*I don't know exactly if Peergos is still considered in alpha or switched to beta, but keep in mind that it might be more experimental that you would like!*
<!--To give ourselves some courage in this adventure, let's start with a nice screenshot of their web UI:
![Peergos Web UI](./peergos.jpg)-->
Starting Peergos on top of Garage required some small patches on both sides,
but in the end, we were able to get it working.
I was able to upload my file, see it in the interface, create a link to share it,
rename it, move it in a folder, and so on:
![A screenshot of the Peergos interface](./upload.png)
At the same time, the fans of my computer started to become a bit loud.
A quick look at Grafana shows that Garage is still very busy.
![Screenshot of a grafana plot showing requests per second over time](./grafa.png)
*Legend: y axis = requests per 10 seconds on log(10) scale, x axis = time*
Again, the workload is dominated by the `HeadObject` requests.
After getting a look at `~/.peergos/.ipfs/config`, it seems the IPFS configuration used by the Peergos project is pretty standard.
And thus, again, we are acting as a DHT server, answering each second to thousands of block requests.
We also some traffic on the `GetObject` endpoints (peaks of ~45 req/sec) and the `OPTIONS` HTTP verb.
This is the traffic generated by Peergos. The `OPTIONS` HTTP verb is here because we use the direct access feature of Peergos: our browser is sending CORS requests to Garage.
Internally, IPFS splits files in blocks of less than 256 kB, my picture is thus split in 2 blocks, requiring 2 requests over Garage to fetch it.
But even by knowing that IPFS split files in small blocks, I can't explain why we have so much `GetObject` requests.
## Try #3: Optimizing IPFS
<!--
Routing = dhtclient
![](./grafa2.png)
-->
We have seen in our 2 precedent tries that the main source of load was the
federation, and more especially, the DHT server. In this section, we want
to artificially remove this problem of the equation by preventing our IPFS node from federating
and see what pressure is put by Peergos alone on our local cluster.
To isolate IPFS, we have set its routing type to `none`, we have cleared its bootstrap node list,
and we configured the swarm socket to listen only on localhost.
Finally, we restart Peergos and observe this more peaceful graph:
![Screenshot of a grafana plot showing requests per second over time](./grafa3.png)
*Legend: y axis = requests per 10 seconds on log(10) scale, x axis = time*
Now, for a given endpoint, we have peaks of around 10 req/sec which is way more reasonable.
Furthermore, we are not hammering anymore our backend with requests on objects that are not here.
The next step would be to gradually allowing back our node to connect to the IPFS network,
while ensuring that the traffic to the S3 cluster remains low. For example, configuring our IPFS
node as a `dhtclient` instead of `dhtserver` would exempt it from answering public DHT requests.
Keeping an in-memory index (as a hashmap and/or blum filter) of the blocks stored on the current node
could also drastically reduce the number of requests.
It could also be interesting to explore ways to run in one process a full IPFS node with a DHT
server on the regularer filesystem datastore, and reserve a second process configured with the S3 datastore
to handle only our Peergos data.
However, with these optimizations, the best we can expect is the traffic we have on the previous plot.
From a theoretical perspective, it is still higher than the optimal number of requests.
On S3, storing a file, downloading a file, listing available files are all actions that can be done in a single request.
Even if all requests have not the same cost on the cluster, processing a request has a non neglictible fixed cost.
## S3 and IPFS are incompatible?
*Text by Alex*
## Conclusion
Running IPFS over a S3 backend does not quite work out of the box in term of performances yet.
We have identified some possible measures for improvement (disabling the DHT server, keeping an in-memory index of the blocks, using the S3 backend only for your data)
that might allow you to still run an IPFS node over Garage.
From a design perspective, it seems however that the numerous small blocks created by IPFS
do not map trivially to efficient S3 requests, and thus could be a limiting factor to any optimization work.
As part of our test journey, we read some posts about performance issues on IPFS (eg. [#6283 - Reduce the impact of the DHT](https://github.com/ipfs/go-ipfs/issues/6283)) that are not
linked with the S3 connector. We might be negatively influenced by our failure to connect IPFS with S3,
but we are tempted to think that in any case, IPFS will be ressource intensive for your hardware.
On our side, we will continue our investigations towards more *minimalist* software that tends to limit the
number of requests they send.
This choice makes sense for us as we want to reduce the ecological impact of our services
by deploying optimized software on a limited number of second-hand servers.
*Yes we are aware of the existence of Nextcloud, Owncloud, Owncloud Infinite Scale, Seafile, Filestash, Pydio, SOLID, Remote Storage, etc.
We might even try one of them in a future post, so stay tuned!*

Binary file not shown.

After

Width:  |  Height:  |  Size: 221 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 110 KiB