| 1 | #include "duckdb_miniparquet.hpp" |
| 2 | |
| 3 | #include "miniparquet.h" |
| 4 | #include "duckdb/function/table_function.hpp" |
| 5 | #include "duckdb/parser/parsed_data/create_table_function_info.hpp" |
| 6 | #include "duckdb/main/client_context.hpp" |
| 7 | #include "duckdb/common/types/date.hpp" |
| 8 | #include "duckdb/common/types/timestamp.hpp" |
| 9 | #include "utf8proc_wrapper.hpp" |
| 10 | #include "duckdb.hpp" |
| 11 | |
| 12 | |
| 13 | using namespace duckdb; |
| 14 | using namespace miniparquet; |
| 15 | |
| 16 | struct ParquetScanFunctionData : public TableFunctionData { |
| 17 | ParquetScanFunctionData(string filename) |
| 18 | : position(0), pf(filename) { |
| 19 | } |
| 20 | vector<SQLType> sql_types; |
| 21 | idx_t position; |
| 22 | ParquetFile pf; |
| 23 | ResultChunk rc; |
| 24 | ScanState s; |
| 25 | }; |
| 26 | |
| 27 | struct ParquetScanFunction : public TableFunction { |
| 28 | ParquetScanFunction() |
| 29 | : TableFunction("parquet_scan" , {SQLType::VARCHAR}, parquet_scan_bind, parquet_scan_function, nullptr){}; |
| 30 | |
| 31 | static unique_ptr<FunctionData> parquet_scan_bind(ClientContext &context, vector<Value> inputs, |
| 32 | vector<SQLType> &return_types, vector<string> &names) { |
| 33 | |
| 34 | auto file_name = inputs[0].GetValue<string>(); |
| 35 | auto res = make_unique<ParquetScanFunctionData>(file_name); |
| 36 | res->pf.initialize_result(res->rc); |
| 37 | |
| 38 | for (auto& col : res->pf.columns) { |
| 39 | names.push_back(col->name); |
| 40 | SQLType type; |
| 41 | switch (col->type) { |
| 42 | case parquet::format::Type::BOOLEAN: |
| 43 | type = SQLType::BOOLEAN; |
| 44 | break; |
| 45 | case parquet::format::Type::INT32: |
| 46 | type = SQLType::INTEGER; |
| 47 | break; |
| 48 | case parquet::format::Type::INT64: |
| 49 | type = SQLType::BIGINT; |
| 50 | break; |
| 51 | case parquet::format::Type::INT96: // always a timestamp? |
| 52 | type = SQLType::TIMESTAMP; |
| 53 | break; |
| 54 | case parquet::format::Type::FLOAT: |
| 55 | type = SQLType::FLOAT; |
| 56 | break; |
| 57 | case parquet::format::Type::DOUBLE: |
| 58 | type = SQLType::DOUBLE; |
| 59 | break; |
| 60 | // case parquet::format::Type::FIXED_LEN_BYTE_ARRAY: { |
| 61 | // todo some decimals yuck |
| 62 | case parquet::format::Type::BYTE_ARRAY: |
| 63 | type = SQLType::VARCHAR; |
| 64 | break; |
| 65 | |
| 66 | default: |
| 67 | throw NotImplementedException("Invalid type" ); |
| 68 | break; |
| 69 | } |
| 70 | return_types.push_back(type); |
| 71 | res->sql_types.push_back(type); |
| 72 | } |
| 73 | |
| 74 | |
| 75 | return move(res); |
| 76 | } |
| 77 | |
| 78 | template <class T> static void scan_parquet_column(ResultColumn& parquet_col, idx_t count, idx_t offset, Vector &out) { |
| 79 | auto src_ptr = (T *)parquet_col.data.ptr; |
| 80 | auto tgt_ptr = FlatVector::GetData<T>(out); |
| 81 | |
| 82 | for (idx_t row = 0; row < count; row++) { |
| 83 | tgt_ptr[row] = src_ptr[row + offset]; |
| 84 | } |
| 85 | } |
| 86 | |
| 87 | // surely they are joking |
| 88 | static constexpr int64_t kJulianToUnixEpochDays = 2440588LL; |
| 89 | static constexpr int64_t kMillisecondsInADay = 86400000LL; |
| 90 | static constexpr int64_t kNanosecondsInADay = kMillisecondsInADay * 1000LL * 1000LL; |
| 91 | |
| 92 | static int64_t impala_timestamp_to_nanoseconds(const Int96 &impala_timestamp) { |
| 93 | int64_t days_since_epoch = impala_timestamp.value[2] |
| 94 | - kJulianToUnixEpochDays; |
| 95 | int64_t nanoseconds = |
| 96 | *(reinterpret_cast<const int64_t*>(&(impala_timestamp.value))); |
| 97 | return days_since_epoch * kNanosecondsInADay + nanoseconds; |
| 98 | } |
| 99 | |
| 100 | |
| 101 | static void parquet_scan_function(ClientContext &context, vector<Value> &input, DataChunk &output, |
| 102 | FunctionData *dataptr) { |
| 103 | auto &data = *((ParquetScanFunctionData *)dataptr); |
| 104 | |
| 105 | if (data.position >= data.rc.nrows) { |
| 106 | if (!data.pf.scan(data.s, data.rc)) { |
| 107 | return; |
| 108 | } |
| 109 | data.position = 0; |
| 110 | } |
| 111 | idx_t this_count = std::min((idx_t)STANDARD_VECTOR_SIZE, data.rc.nrows - data.position); |
| 112 | assert(this_count > 0); |
| 113 | output.SetCardinality(this_count); |
| 114 | |
| 115 | for (idx_t col_idx = 0; col_idx < output.column_count(); col_idx++) { |
| 116 | auto& col = data.rc.cols[col_idx]; |
| 117 | |
| 118 | for (idx_t row = 0; row < this_count; row++) { |
| 119 | FlatVector::SetNull(output.data[col_idx], row, !((uint8_t*) col.defined.ptr)[row + data.position]); |
| 120 | } |
| 121 | |
| 122 | switch (data.sql_types[col_idx].id) { |
| 123 | case SQLTypeId::BOOLEAN: |
| 124 | scan_parquet_column<bool>(col, this_count, data.position, output.data[col_idx]); |
| 125 | break; |
| 126 | case SQLTypeId::INTEGER: |
| 127 | scan_parquet_column<int32_t>(col, this_count, data.position, output.data[col_idx]); |
| 128 | break; |
| 129 | case SQLTypeId::BIGINT: |
| 130 | scan_parquet_column<int64_t>(col, this_count, data.position, output.data[col_idx]); |
| 131 | break; |
| 132 | case SQLTypeId::FLOAT: |
| 133 | scan_parquet_column<float>(col, this_count, data.position, output.data[col_idx]); |
| 134 | break; |
| 135 | case SQLTypeId::DOUBLE: |
| 136 | scan_parquet_column<double>(col, this_count, data.position, output.data[col_idx]); |
| 137 | break; |
| 138 | case SQLTypeId::TIMESTAMP: { |
| 139 | auto tgt_ptr = (timestamp_t *)FlatVector::GetData(output.data[col_idx]); |
| 140 | for (idx_t row = 0; row < this_count; row++) { |
| 141 | auto impala_ns = impala_timestamp_to_nanoseconds(((Int96*) col.data.ptr)[row + data.position]); |
| 142 | |
| 143 | auto ms = impala_ns / 1000000; // nanoseconds |
| 144 | auto ms_per_day = (int64_t)60 * 60 * 24 * 1000; |
| 145 | date_t date = Date::EpochToDate(ms / 1000); |
| 146 | dtime_t time = (dtime_t)(ms % ms_per_day); |
| 147 | tgt_ptr[row] = Timestamp::FromDatetime(date, time); |
| 148 | ; |
| 149 | } |
| 150 | // ugh |
| 151 | break; |
| 152 | } break; |
| 153 | case SQLTypeId::VARCHAR: { |
| 154 | auto src_ptr = (char**)col.data.ptr; |
| 155 | auto tgt_ptr = (string_t *)FlatVector::GetData(output.data[col_idx]); |
| 156 | |
| 157 | for (idx_t row = 0; row < this_count; row++) { |
| 158 | auto val = src_ptr[row + data.position]; |
| 159 | if (!FlatVector::IsNull(output.data[col_idx], row)) { |
| 160 | auto utf_type = Utf8Proc::Analyze(val); |
| 161 | if (utf_type == UnicodeType::INVALID) { |
| 162 | throw Exception("Invalid UTF in Parquet file" ); |
| 163 | } else if (utf_type == UnicodeType::ASCII) { |
| 164 | tgt_ptr[row] = StringVector::AddString(output.data[col_idx], val); |
| 165 | } else { |
| 166 | auto val_norm = Utf8Proc::Normalize(val); |
| 167 | tgt_ptr[row] = StringVector::AddString(output.data[col_idx], val_norm); |
| 168 | free(val_norm); |
| 169 | } |
| 170 | } |
| 171 | } |
| 172 | break; |
| 173 | } |
| 174 | default: |
| 175 | throw NotImplementedException("Unsupported type " + SQLTypeToString(data.sql_types[col_idx])); |
| 176 | } |
| 177 | } |
| 178 | data.position += this_count; |
| 179 | } |
| 180 | }; |
| 181 | |
| 182 | void Parquet::Init(DuckDB& db) { |
| 183 | ParquetScanFunction scan_fun; |
| 184 | CreateTableFunctionInfo info(scan_fun); |
| 185 | |
| 186 | Connection conn(db); |
| 187 | conn.context->transaction.BeginTransaction(); |
| 188 | conn.context->catalog.CreateTableFunction(*conn.context, &info); |
| 189 | conn.context->transaction.Commit(); |
| 190 | } |
| 191 | |