[C++][Parquet] Minor: Declare lifetime of shared_ptr<Page> on PageReader

[mapleFU]

Describe the enhancement requested

In C++ Parquet, SerializedPageReader holds a decryption_buffer_ and decompression_buffer_, and when NextPage called, it will reuse that buffer. So, Page's buffer data is bounded on SerializedPageReader::NextPage, when NextPage is called, it's data might be reset.

I have two problems here:

1. If it's by design, why don't we cache raw data in SerializePageReader?
2. Should we note these in comments?

Component(s)

C++, Parquet

[mapleFU]

@pitrou @wjones127 mind take a look?

[cyb70289]

It might be easier to review if you can paster the code link here.

[mapleFU]

The interface:

// Abstract page iterator interface. This way, we can feed column pages to the
// ColumnReader through whatever mechanism we choose
class PARQUET_EXPORT PageReader {
  using DataPageFilter = std::function<bool(const DataPageStats&)>;

 public:
  virtual ~PageReader() = default;

  // @returns: shared_ptr<Page>(nullptr) on EOS, std::shared_ptr<Page>
  // containing new Page otherwise
  virtual std::shared_ptr<Page> NextPage() = 0;

The actual:

// This subclass delimits pages appearing in a serialized stream, each preceded
// by a serialized Thrift format::PageHeader indicating the type of each page
// and the page metadata.
class SerializedPageReader : public PageReader {
 public:
  SerializedPageReader(std::shared_ptr<ArrowInputStream> stream, int64_t total_num_values,
                       Compression::type codec, const ReaderProperties& properties,
                       const CryptoContext* crypto_ctx, bool always_compressed)
      : properties_(properties),
        stream_(std::move(stream)),
        decompression_buffer_(AllocateBuffer(properties_.memory_pool(), 0)),
        decryption_buffer_(AllocateBuffer(properties_.memory_pool(), 0)) {}

It (implicitly) depend on SerializedPageReader's buffer