libstore/binary-cache: default to zstd for compression

libarchive's xz offers single threaded xz compression which is very slow
and provides ~10-20Mbps compression speed in addition to maxing a core.

In exchange, it achieves optimal compression ratios among all our
compression methods.

Nonetheless, xz prevent the saturation of 1Gbps+ connections and slow
down significantly decompression for end users. As these connections and
faster hardware is becoming prevalent for cache servers and clients, we
offer to default to zstd.

Lix is a "compress once, decompress many times" application. To avoid
incurring a high penalty to end users very sensitive to compress ratio
(very slow Internet connections), we dampen the consequences of
switching to zstd by increasing the default zstd level to 12.

On one example, xz will compress a 4.4GB file to 632MB, zstd on 12 will
compress it to 775MB, that is a ~18 % increase over the optimal xz
compression. zstd took 18 seconds to produce this file.

Increasing to level 14 leads to a 773MB file while taking 37s.
Increasing to level 16 leads to 735MB file while taking 66s.

Finally, xz took 77s, so a 50 % reduction in time taken to compress in
exchange of an increase of 18 % of the compressed size.

This change will reduce issues encountered in #945 but is probably not
the root cause.

References:

- https://discourse.nixos.org/t/switch-cache-nixos-org-to-zstd-to-fix-slow-nixos-updates-nix-downloads/23961

Change-Id: I7beda2bf2c1fed146dcb797b8f85dc290c486ab2
Signed-off-by: Raito Bezarius <raito@lix.systems>

doc/manual/rl-next/enable-zstd-for-binary-cache.md

@@ -0,0 +1,36 @@
+---
+synopsis: Enable high compress ratio zstd compression by default for binary caches uploads
+issues: [fj#945]
+cls: [4503]
+category: "Breaking Changes"
+credits: [horrors, raito]
+---
+
+The default compression method for binary cache uploads has been switched from
+[`xz`](https://github.com/tukaani-project/xz) to
+[`zstd`](https://github.com/facebook/zstd) to address performance and usability
+issues related to modern hardware and high-speed connections.
+
+## Why?
+
+`xz` offers compression ratios but is single-threaded in our implementation and
+very slow (~10-20 Mbps in our test), preventing full utilization of 100Mbps+
+connections and significantly slowing decompression for end users.
+
+Lix is a "compress once, decompress many" application: build farms can afford
+to spend more time compressing to achieve a faster download transfer for the
+end user. More importantly, it matters that all end users spend the least
+amount of time decompressing.
+
+## What about compression ratios?
+
+`zstd` cannot achieve the same peaks as `xz`, nonetheless, `zstd` compression
+level has been increased to level 12 by default to balance compression ratio
+and performance.
+
+## Synthetic test case data
+
+* **xz** (default compression level) on a 4.4GB file: ~632MB (77s)
+* **zstd** (level 12) on the same file: ~775MB (18s), 18% larger but 50% faster
+* **zstd** (level 14): ~773MB (37s)
+* **zstd** (level 16): ~735MB (66s)

lix/libstore/binary-cache-store.hh

@@ -19,8 +19,12 @@ struct BinaryCacheStoreConfig : virtual StoreConfig
 {
     using StoreConfig::StoreConfig;
 
-    const Setting<std::string> compression{this, "xz", "compression",
-        "NAR compression method (`xz`, `bzip2`, `gzip`, `zstd`, or `none`)."};
+    const Setting<std::string> compression{
+        this,
+        "zstd",
+        "compression",
+        "NAR compression method (`xz`, `bzip2`, `gzip`, `zstd`, or `none`)."
+    };
 
     const Setting<bool> writeNARListing{this, false, "write-nar-listing",
         "Whether to write a JSON file that lists the files in each NAR."};
@@ -40,12 +44,19 @@ struct BinaryCacheStoreConfig : virtual StoreConfig
     const Setting<bool> parallelCompression{this, false, "parallel-compression",
         "Enable multi-threaded compression of NARs. This is currently only available for `xz` and `zstd`."};
 
-    const Setting<int> compressionLevel{this, -1, "compression-level",
+    const Setting<int> compressionLevel{
+        this,
+        -1,
+        "compression-level",
         R"(
           The *preset level* to be used when compressing NARs.
           The meaning and accepted values depend on the compression method selected.
           `-1` specifies that the default compression level should be used.
-        )"};
+
+          Note: when using zstd `-1` will select level 12 to approximately match xz compression
+          ratios at default settings rather than the zstd library default of 3.
+        )"
+    };
 };
 
 

lix/libutil/compression.cc

@@ -465,6 +465,17 @@ ref<CompressionSink> makeCompressionSink(const std::string & method, Sink & next
     std::vector<std::string> la_supports = {
         "bzip2", "compress", "grzip", "gzip", "lrzip", "lz4", "lzip", "lzma", "lzop", "xz", "zstd"
     };
+
+    // NOTE: Lix overrides the default here because we want the default zstd behavior
+    // to perform well on compression ratios in the hopes to approach what xz provided in the past.
+    // We choose one that is much faster than xz while being in range of xz compression ratios.
+    //
+    // In our experience, further levels provides marginal benefits but makes the compression speed
+    // much slower in exchange.
+    if (level == COMPRESSION_LEVEL_DEFAULT && method == "zstd") {
+        level = 12;
+    }
+
     if (std::find(la_supports.begin(), la_supports.end(), method) != la_supports.end()) {
         return make_ref<ArchiveCompressionSink>(nextSink, method, parallel, level);
     }

package.nix

@@ -61,6 +61,7 @@
   util-linuxMinimal ? utillinuxMinimal,
   utillinuxMinimal ? null,
   xz,
+  zstd,
   yq,
 
   busybox-sandbox-shell,
@@ -331,6 +332,7 @@ stdenv.mkDerivation (finalAttrs: {
     jq
     yq
     lsof
+    zstd
   ]
   ++ lib.optional hostPlatform.isLinux util-linuxMinimal
   ++ lib.optional (!officialRelease && buildUnreleasedNotes) build-release-notes

tests/functional/binary-cache.sh

@@ -73,10 +73,10 @@ basicDownloadTests
 # Test whether Lix notices if the NAR doesn't match the hash in the NAR info.
 clearStore
 
-nar=$(ls $cacheDir/nar/*.nar.xz | head -n1)
+nar=$(ls $cacheDir/nar/*.nar.zst | head -n1)
 mv $nar $nar.good
 mkdir -p $TEST_ROOT/empty
-nix-store --dump $TEST_ROOT/empty | xz > $nar
+nix-store --dump $TEST_ROOT/empty | zstd - --stdout > $nar
 
 expect 1 nix-build --substituters "file://$cacheDir" --no-require-sigs dependencies.nix -o $TEST_ROOT/result 2>&1 | tee $TEST_ROOT/log
 grepQuiet "hash mismatch" $TEST_ROOT/log

tests/functional/substitute-truncated-nar.sh

@@ -15,13 +15,13 @@ path=$(build)
 nix copy --to "$BINARY_CACHE" "$path"
 nix-collect-garbage >/dev/null 2>&1
 
-nar=0bylmx35yjy2b1b4k7gjsl7i4vc03cpmryb41grfb1mp40n3hifl.nar.xz
+nar=0c3y7p42issm0ydjilwvk0drv958p4p4d2d6c7y5ksmzmbf7rfhg.nar.zst
 
 [ -e $cacheDir/nar/$nar ] || fail "long nar missing?"
 
-xzcat $cacheDir/nar/$nar > $TEST_HOME/tmp
+zstdcat $cacheDir/nar/$nar > $TEST_HOME/tmp
 truncate -s $(( $(stat -c %s $TEST_HOME/tmp) - 10 )) $TEST_HOME/tmp
-xz < $TEST_HOME/tmp > $cacheDir/nar/$nar
+zstd - --stdout < $TEST_HOME/tmp > $cacheDir/nar/$nar
 
 # Copying back '$path' from the binary cache. This should fail as it is truncated
 if build --option substituters "$BINARY_CACHE" --option require-sigs false -j0; then