PERF: Skip header check in parquet reader (rapidsai#22679)

The parquet reader previously required three reads to read the parquet footer:

1. A 4 byte read to check the header for the parquet magic bytes
2. An 8 byte read to read the footer length and footer parquet magic bytes
3. A varaible-length read for the footer metadata

We don't really care about ensuring that the header is valid. For high-latency storage, it's not worth the extra read.

Part of rapidsai#22668, which also proposes to remove the second 8-byte read. But this is a smaller change that should be less controversial.

Authors:
  - Tom Augspurger (https://github.com/TomAugspurger)

Approvers:
  - Bradley Dice (https://github.com/bdice)
  - Vukasin Milovanovic (https://github.com/vuule)

URL: rapidsai#22679

Author: TomAugspurger

cpp/src/io/parquet/io_utils/parquet_io_utils.cpp

@@ -37,12 +37,10 @@ std::unique_ptr<cudf::io::datasource::buffer> fetch_footer_to_host(cudf::io::dat
   constexpr auto ender_len  = sizeof(file_ender_s);
   size_t const len          = datasource.size();
 
-  auto header_buffer = datasource.host_read(0, header_len);
-  auto const header  = reinterpret_cast<file_header_s const*>(header_buffer->data());
-  auto ender_buffer  = datasource.host_read(len - ender_len, ender_len);
-  auto const ender   = reinterpret_cast<file_ender_s const*>(ender_buffer->data());
   CUDF_EXPECTS(len > header_len + ender_len, "Incorrect data source");
-  CUDF_EXPECTS(header->magic == detail::parquet_magic, "Corrupted header");
+
+  auto ender_buffer = datasource.host_read(len - ender_len, ender_len);
+  auto const ender  = reinterpret_cast<file_ender_s const*>(ender_buffer->data());
   CUDF_EXPECTS(ender->magic == detail::parquet_magic, "Corrupted footer");
   CUDF_EXPECTS(ender->footer_len != 0 && ender->footer_len <= (len - header_len - ender_len),
                "Incorrect footer length");

cpp/tests/io/parquet_reader_test.cpp

@@ -4375,6 +4375,27 @@ TEST_F(ParquetReaderTest, LateBindSourceInfo)
   CUDF_TEST_EXPECT_TABLES_EQUAL(result.tbl->view(), expected->view());
 }
 
+TEST_F(ParquetReaderTest, InvalidFooterMagic)
+{
+  auto const expected = create_random_fixed_table<int>(4, 4, false);
+
+  std::vector<char> buffer;
+  cudf::io::write_parquet(
+    cudf::io::parquet_writer_options::builder(cudf::io::sink_info{&buffer}, *expected));
+
+  constexpr std::array<char, 4> bad_magic{'B', 'A', 'D', '!'};
+  ASSERT_GE(buffer.size(), bad_magic.size());
+  for (size_t i = 0; i < bad_magic.size(); ++i) {
+    buffer[buffer.size() - bad_magic.size() + i] = bad_magic[i];
+  }
+
+  auto const read_opts = cudf::io::parquet_reader_options::builder(
+                           cudf::io::source_info{cudf::host_span<std::byte const>{
+                             reinterpret_cast<std::byte const*>(buffer.data()), buffer.size()}})
+                           .build();
+  EXPECT_THROW(cudf::io::read_parquet(read_opts), cudf::logic_error);
+}
+
 TEST_F(ParquetReaderTest, DecimalTypeOption)
 {
   auto const data = std::vector<int32_t>{1000, 2000, 3000, 4000, 5000};