benchmarks: Add Accept-Encoding: gzip headers (NO REGRESSION)

PR-URL: https://github.com/hasura/graphql-engine-mono/pull/5622
GitOrigin-RevId: 431f7f382573737af9b85966a58c5b262f8cb377

server/benchmarks/README.md

@@ -83,6 +83,11 @@ be ignored.
     https://well-typed.com/blog/2021/01/fragmentation-deeper-look/
     https://well-typed.com/blog/2021/03/memory-return/
 
+- If optimizing or tuning the output/compression codepath:
+  - `chinook`.`*_small_result` and `simple_query_*` queries are average wrt
+    response body size (according to cloud data)
+  - ...and `chinook`.`full_introspection` is ~P95
+
 ## Adding a new benchmark and reviewing
 
 You'll create a new directory under `benchmark_sets/`, and in general can

server/benchmarks/benchmark_sets/chinook/config.query.yaml

@@ -5,6 +5,11 @@ headers:
   X-Hasura-Admin-Secret: my-secret
   X-Hasura-Role: employee
   X-Hasura-User-Id: 4
+  # On cloud about 95% of requests incoming Accept gzip (which as of writing
+  # means we unconditionally compress, which has significant performance
+  # impact), so it's important that we include it in all benchmarks unless we
+  # have an intentional reason not to.
+  Accept-Encoding: gzip
 
 # Anchors to help us DRY below; settings here may be overridden selectively
 constants:
@@ -36,6 +41,12 @@ queries:
   ############################################################################
   # single-table query, small result size; makes use of permissions for
   # filtering; low RPS
+  #
+  # NOTE: According to data from cloud this is a pretty average response (we see
+  # P50 compressed response body size of 123 bytes, and P75 290 bytes):
+  #  
+  #   uncompressed body size: ~600 bytes
+  #   compressed body size:   ~200 bytes   (as of time of writing)
   - name: simple_query_low_load
     <<: *k6_custom
     options:
@@ -79,6 +90,10 @@ queries:
 
 
   ######## Small result size
+  # 
+  #   uncompressed body size: ~650 bytes
+  #   compressed body size:   ~200 bytes
+
   - name: complex_query_low_load_small_result
     <<: *k6_custom
     options:
@@ -130,6 +145,13 @@ queries:
 
 
   ######## Large result size
+  #
+  # NOTE: According to data from cloud, this is somewhere between 
+  #       P90 (2 kB compressed) and P95 (4 kB compressed):
+  #
+  #   uncompressed body size: ~33 kB
+  #   compressed body size:    ~3 kB   (as of time of writing)
+
   - name: complex_query_low_load_large_result
     <<: *k6_custom
     options:
@@ -164,6 +186,12 @@ queries:
   # The standard introspection query from server/src-rsr/introspection.json
   # We don't expect a server to be hammered with these, but they are crucial
   # for clients (like graphiql):
+  #
+  # NOTE: According to data from cloud, this is somewhere between 
+  # P95 (4 kB compressed) and P99 (40 kB compressed):
+  #
+  #   uncompressed body size: ~190 kB
+  #   compressed body size:    ~13 kB   (as of time of writing)
 
   - name: full_introspection
     <<: *k6_custom

server/benchmarks/benchmark_sets/huge_schema/config.query.yaml

@@ -1,8 +1,8 @@
 # This tells graphql-bench that it's testing a hasura instance and should
 # collect some additional metrics:
 extended_hasura_checks: true
-# headers:
-#   X-Hasura-Admin-Secret: my-secret
+headers:
+  Accept-Encoding: gzip
 
 # Anchors to help us DRY below; settings here may be overridden selectively
 constants:

server/benchmarks/benchmark_sets/remote_schema/config.query.yaml

@@ -1,8 +1,8 @@
 # This tells graphql-bench that it's testing a hasura instance and should
 # collect some additional metrics:
 extended_hasura_checks: true
-# headers:
-#   X-Hasura-Admin-Secret: my-secret
+headers:
+  Accept-Encoding: gzip
 
 # Anchors to help us DRY below; settings here may be overridden selectively
 constants:

server/src-lib/Hasura/Server/Compression.hs

@@ -19,21 +19,26 @@ data CompressionType
 compressionTypeToTxt :: CompressionType -> Text
 compressionTypeToTxt CTGZip = "gzip"
 
+-- | Maybe compress the response body
 compressResponse ::
   NH.RequestHeaders ->
   BL.ByteString ->
   (BL.ByteString, Maybe NH.Header, Maybe CompressionType)
 compressResponse reqHeaders unCompressedResp =
-  let compressionTypeM = getRequestedCompression reqHeaders
+  let compressionTypeM = getAcceptedCompression reqHeaders
       appendCompressionType (res, headerM) = (res, headerM, compressionTypeM)
       gzipCompressionParams =
+        -- See Note [Compression ratios]
         GZ.defaultCompressParams {GZ.compressLevel = GZ.compressionLevel 1}
    in appendCompressionType $ case compressionTypeM of
         Just CTGZip -> (GZ.compressWith gzipCompressionParams unCompressedResp, Just gzipHeader)
         Nothing -> (unCompressedResp, Nothing)
 
-getRequestedCompression :: NH.RequestHeaders -> Maybe CompressionType
-getRequestedCompression reqHeaders
+-- | Which, if any, compressed encodings can the client accept?
+--
+-- https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers/Accept-Encoding
+getAcceptedCompression :: NH.RequestHeaders -> Maybe CompressionType
+getAcceptedCompression reqHeaders
   | "gzip" `elem` acceptEncodingVals = Just CTGZip
   | otherwise = Nothing
   where
@@ -41,3 +46,31 @@ getRequestedCompression reqHeaders
       concatMap (splitHeaderVal . snd) $
         filter (\h -> fst h == NH.hAcceptEncoding) reqHeaders
     splitHeaderVal bs = map T.strip $ T.splitOn "," $ bsToTxt bs
+
+{-
+Note [Compression ratios]
+~~~~~~~~~~~~~~~~~~~~~~~~~
+
+I did some measurements of compression ratios at `gzip -1` (libc) of some
+randomly generated json, real json datasets, and output from our benchmarked
+chinook queries:
+
+    2552/6131    = 0.41
+    4666/8718    = 0.53
+    13921/27131  = 0.51
+    5895/8879    = 0.66  <----- completely random strings
+    8634/28261   = 0.30
+    70422/372466 = 0.18
+
+    200/600      = 0.33  <----| from chinook graphql benchmarks
+    3000/33000   = 0.09  <----|
+    13000/190000 = 0.07  <----'
+
+Given these numbers I would suggest using a rule-of-thumb expected compression
+ratio between 2:1 and 10:1, depending on what being conservative means in the
+context.
+
+I didn't test higher compression levels much, but `gzip -4` for the most part
+resulted in less than 10% smaller output on random json, and ~30% on our highly
+compressible benchmark output.
+-}