fixed chunk size #173

philandstuff · 2024-03-04T17:59:05Z

We have a fair amount of complexity (EqualSplit etc) involved in dynamically splitting files down into chunks based on available concurrency.

We could be less clever and use a fixed chunk size, with a fixed buffer size to go with it. This would dovetail nicely with a sync.Pool of fixed-size buffers, and also with a configuration option to limit the total allocated RAM (implemented as a semaphore guarding how many buffers can be fetched from the sync.Pool).

This commit only does consistent hashing mode; we should do buffer mode as well if we do this work.

We have a fair amount of complexity (EqualSplit etc) involved in dynamically splitting files down into chunks based on available concurrency. We could be less clever and use a fixed chunk size, with a fixed buffer size to go with it. This would dovetail nicely with a sync.Pool of fixed-size buffers, and also with a configuration option to limit the total allocated RAM (implemented as a semaphore guarding how many buffers can be fetched from the sync.Pool). This commit only does consistent hashing mode; we should do buffer mode as well if we do this work.

dkhokhlov · 2024-03-04T18:43:29Z

side note:
https://github.com/replicate/pget/pull/173/files#diff-ec1a0e746b82d63c33a617e0b8dd205d92ac3cec2213443a6e29af27aaa37f28R139-R145
checks for one var (firstReqResult.err) but uses another var (err) in diag.

For const and var shared between buffer and consistent-hash mode move these constants to a dedicated .go file. These are two distinct downloaders and neither should be responsible for explicitly shared elements.

Wire up the --chunk-size option and deprecate --minimum-chunk-size.

The concurrency calculation is no longer used in consistent-hashing. Remove it.

Update Buffer downloader to use ChunkSize instead of MinChunkSize. Options have been modified to only contain ChunkSize now as the root and multifile commands no longer reference MinChunkSize. This change is in support of moving towards simplifying the chunksize logic. Additionally this will make reuse of buffers significantly easier as the next phase.

pkg/download/buffer.go

If chunkSize doesn't divide contentSize evenly, and chunkSize < 0.5*contentSize, we try to download a chunk bigger than chunkSize. This should not be possible.

Every chunk except the last is now the same chunkSize, which is the intention of the code. We no longer dynamically change chunk size. Also, simplify the consistent hashing code to be more similar to buffer mode, and remove the dead EqualSplit code.

README.md

philandstuff · 2024-03-05T12:15:59Z

cmd/root/root.go

@@ -199,7 +202,7 @@ func runRootCMD(cmd *cobra.Command, args []string) error {
 // rootExecute is the main function of the program and encapsulates the general logic
 // returns any/all errors to the caller.
 func rootExecute(ctx context.Context, urlString, dest string) error {
-	minChunkSize, err := humanize.ParseBytes(viper.GetString(config.OptMinimumChunkSize))
+	chunkSize, err := humanize.ParseBytes(viper.GetString(config.OptMinimumChunkSize))


Suggested change

chunkSize, err := humanize.ParseBytes(viper.GetString(config.OptMinimumChunkSize))

chunkSize, err := humanize.ParseBytes(viper.GetString(config.OptChunkSize))

We should still check MinimumChunkSize and adjust ChunkSize upwards to satisfy it if it's set.

I'm not clear why we are bothering to keep min chunk size. It seems... odd? Let's just specify the chunk size we want. Having two controls for the same thing is how we got into the mess of too many ways to control concurrency.

If someone wants a chunk size of 1byte... it's silly but it's fine.

because pget is already deployed to many models which set it

if they set min chunk size to 500MiB we should respect that and not have a 125MiB chunk size.

I had wired up minchunk size and chunk size to the same value behind the scenes so they should be able to be set either flag but with --minchunk size deprecated.

The exception is in prod where our value would continue to override (since we set with the config map).

The only caveat is if you specified both, the last one processed wins.

I see that I may not have been clear about that change. Totally on me and I'll be better call those things out in the future.

Oops, yes I missed that. However:

we should not refer to the deprecated name on this line, so my suggestion above still stands

we wire both args up to the chunkSize variable but we never read that variable so it's not obvious to me that it matters (instead we fetch with viper). I don't have a good mental model for what happens? I'm not sure why we bind a variable if we're never reading it?

pflag has docs on how to alias flag names, maybe we should do that instead?

The aliasing is a bit weird with viper involved. Generally using an XXXVarP type instead of XXXP in cobra is the easiest way to alias as it backs to a pointer. I'm really ok either way. Your suggestions do absolutely apply in either case.

why do we have viper, again? is it so we can configure via environment variable? it's all a bit magic, I don't understand it, and i don't find the documentation helpful in the detail (like this variable aliasing thing, I don't think the documentation defines the behaviour here).

Viper is for the ENV binding to the flags. It's a lot of magic

Variable aliasing results in the last processed element winning i'll mkae sure the arguments cannot be both used. It should address the concerns. The ENV bits is a bit more work I'll make sure it's also clear (or at least documented) behavior

pkg/download/options.go

Aliases in Cobra and Viper are relatively complex and viper.RegisterAlias does not do what would be expected as ENV handling is a very special case. This commit fixes up the handling of aliases specifically for chunk size and minimuim chunk size. * CLI Option Set * NEW ENV VAR * Deprecated ENV Var * Default Value For ChunkSize and MinimumChunkSize we now explicitly disallow setting both --chunk-size and --minimum-chunk-size as CLI options. If PGET_CHUNK_SIZE is set, we utilize that value except if the cli option(s) are set. If PGET_MINIMUM_CHUNK_SIZE is set * If PGET_CHUNK_SIZE is set to a non-default value a warning is emitted, and PGET_CHUNK_SIZE is used.. * If PGET_CHUNK_SIZE is the default value, PGET_MINIMUM_CHUNK_SIZE is set as PGET_CHUNK_SIZE and a deprecation warning is emitted. Additionally, viper calls now all reference config.OptChunkSize

tempusfrangit

Approving at this point, however, @philandstuff I'd like you to also take another look at this before we land it.

philandstuff · 2024-03-06T11:49:23Z

cmd/root/root.go

+	changedChunk := cmd.PersistentFlags().Changed(config.OptChunkSize)
+	if changedMin && changedChunk {
+		return fmt.Errorf("--minimum-chunk-size and --chunk-size cannot be used at the same time, use --chunk-size instead")
+	} else if !(changedMin && changedChunk) {


Suggested change

} else if !(changedMin && changedChunk) {

} else {

in this house, we accept the principal of the excluded middle, so this condition is unnecessary.

in fact, because the if block returns, we don't need an else at all

cmd/root/root.go

Simplify conditional. Also, if both PGET_MINIMUM_CHUNK_SIZE and PGET_CHUNK_SIZE are set to the same value, don't emit a warning.

philandstuff requested a review from a team March 4, 2024 17:59

tempusfrangit added 4 commits March 4, 2024 11:50

Move constants in downloads to common.go

9862405

For const and var shared between buffer and consistent-hash mode move these constants to a dedicated .go file. These are two distinct downloaders and neither should be responsible for explicitly shared elements.

Wire up --chunk-size to cli options

6f9c682

Wire up the --chunk-size option and deprecate --minimum-chunk-size.

Cleanup Linting

d700109

The concurrency calculation is no longer used in consistent-hashing. Remove it.

philandstuff commented Mar 5, 2024

View reviewed changes

pkg/download/buffer.go Outdated Show resolved Hide resolved

Failing test case

31bc26d

If chunkSize doesn't divide contentSize evenly, and chunkSize < 0.5*contentSize, we try to download a chunk bigger than chunkSize. This should not be possible.

philandstuff force-pushed the fixed-chunk-size branch from 3254d15 to 31bc26d Compare March 5, 2024 11:58

Fix buffermode

7ec1a9b

Every chunk except the last is now the same chunkSize, which is the intention of the code. We no longer dynamically change chunk size. Also, simplify the consistent hashing code to be more similar to buffer mode, and remove the dead EqualSplit code.

philandstuff commented Mar 5, 2024

View reviewed changes

README.md Outdated Show resolved Hide resolved

philandstuff commented Mar 5, 2024

View reviewed changes

README.md Outdated Show resolved Hide resolved

philandstuff commented Mar 5, 2024

View reviewed changes

pkg/download/options.go Outdated Show resolved Hide resolved

philandstuff and others added 5 commits March 5, 2024 12:18

Update pkg/download/options.go

870dcd8

Update README.md

6517274

Update README.md

401e975

continue to respect minchunksize

fae7259

tempusfrangit force-pushed the fixed-chunk-size branch from 716bf17 to e71b6b2 Compare March 5, 2024 23:11

tempusfrangit changed the title ~~WIP: fixed chunk size~~ fixed chunk size Mar 5, 2024

tempusfrangit approved these changes Mar 5, 2024

View reviewed changes

philandstuff commented Mar 6, 2024

View reviewed changes

cmd/root/root.go Show resolved Hide resolved

more aliasing cleanup

a41a4c4

Simplify conditional. Also, if both PGET_MINIMUM_CHUNK_SIZE and PGET_CHUNK_SIZE are set to the same value, don't emit a warning.

philandstuff merged commit 58cccd0 into main Mar 6, 2024
5 checks passed

philandstuff deleted the fixed-chunk-size branch March 6, 2024 14:48

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fixed chunk size #173

fixed chunk size #173

philandstuff commented Mar 4, 2024

dkhokhlov commented Mar 4, 2024

philandstuff Mar 5, 2024

tempusfrangit Mar 5, 2024

philandstuff Mar 5, 2024

tempusfrangit Mar 5, 2024 •

edited

Loading

philandstuff Mar 5, 2024

tempusfrangit Mar 5, 2024

philandstuff Mar 5, 2024

tempusfrangit Mar 5, 2024

tempusfrangit left a comment

philandstuff Mar 6, 2024

philandstuff Mar 6, 2024

	chunkSize, err := humanize.ParseBytes(viper.GetString(config.OptMinimumChunkSize))
	chunkSize, err := humanize.ParseBytes(viper.GetString(config.OptChunkSize))

fixed chunk size #173

fixed chunk size #173

Conversation

philandstuff commented Mar 4, 2024

dkhokhlov commented Mar 4, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tempusfrangit Mar 5, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tempusfrangit left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tempusfrangit Mar 5, 2024 •

edited

Loading