Error correction codes are quite magical. Instead of splitting in 100 pieces each piece 3 times, you split it in 300 distinct pieces of size 1/100 * filesize, which any 100 can reconstruct the file. So basically you can drop 2/3 of the pieces and yet be able to reconstruct the file.

So as long as your nodes uptime is greater than 33% + eps, the file is not gone. The eps is quite small because of the law of great number.

For example if your node uptime is 50%, the pieces will be in average available on 300 * 50% = 150 +- C*uptimevariance /sqrt(N) nodes which is greater than 100 with probability ~ 1, so the file will be recoverable almost surely.