Recently ran some data on the test network and felt that the current storage track has really been skewed by established projects. Take Arweave for example, which focuses on permanent storage, it is indeed appealing, but the one-time buyout Endowment model actually suppresses the capital utilization rate, especially when you want to store large files that are not financially oriented, the cost directly deters you. I've been playing around with Walrus these days and found that their two-dimensional erasure code is indeed quite interesting; unlike Filecoin, which requires extremely expensive hardware for packaging proofs, it makes the data sharding step particularly lightweight.
In practice, I uploaded a video file of several hundred megabytes, and the speed was significantly faster than pinning on IPFS. The logic behind this is actually leveraging the high concurrency characteristics of the Sui network, specifically for handling large Blob data. But this doesn't mean it's perfect; I found that the error rate when using the CLI tool is not low, especially when the number of concurrent requests increases, the node response has noticeable delays, which is clearly a result of insufficient early code optimization. Compared to Celestia, Walrus clearly focuses more on scenarios where 'even non-transactional data must maintain high availability', rather than purely a data availability layer.
I think the most interesting point is how it handles the incentives for storage nodes. From the current version, it seems to want to exchange lower hardware thresholds for a broader node distribution, which is the right approach. If it can really reduce storage costs to the level of Web2 cloud services, that would be a real achievement. However, the documentation is truly difficult to describe; many parameters are explained ambiguously, causing me to debug half the night. I hope these issues can be resolved before the mainnet launch; after all, the storage track is not short of stories, but what it lacks is infrastructure that can not only store but also be cheap to retrieve.
At first, it's indeed easy to get confused with this thing, but once you understand its sharding logic, you will find that this may be the most suitable storage solution for high-frequency interactive DApps at the moment.
