The table block. A table is always the whole value of a data packet, sent by pointer with format Arrow (Packet protocol), even when shared memory is otherwise disabled. Tables cannot be nested inside other values. The block holds a header, then node descriptors, buffer descriptors, a string table, and the data buffers, each buffer aligned to 64 bytes from the start of the block, with zeroed padding.

A node is one Arrow array: the root is a struct whose children are the columns, and a nested column has nodes of its own. Each 64-byte node descriptor carries what the Arrow C Data Interface needs to rebuild the array: length, null count, flags, the format string and column name (as offsets into the string table), and the index and count of its buffers and children. Each 32-byte buffer descriptor gives a buffer’s offset and size within the block. Slices are compacted and dictionaries decoded before a table is written, so every node starts at offset 0 and has no dictionary.

Zero-copy import. A receiver validates the block (every offset in bounds, every buffer long enough for its node), then builds Arrow C Data Interface structures whose buffers point into the block, holding a reference on it for as long as the table lives. Python receives a pyarrow record batch, R an arrow record batch, C++ a nanoarrow-based handle, and Rust an arrow-rs record batch, each without copying the data. Converting further, to a pandas or R data frame, is the language library’s own copy. Sending a table writes a new block, except that a pool returning a table it received, unchanged, passes the same block on.

Declared columns. A table’s type lists columns (T: in its schema; Voidstar format and schema strings), and that list is a lower bound: a table may have more columns, which travel after the declared ones. A declared column may be a boolean, integer, float or string, or an optional of one of these; any Arrow type may appear among the extra columns. On send, a declared column of another type is cast if the cast loses nothing, and refused otherwise. On receive, every declared column must be present with an acceptable type, and a column not declared optional must have no nulls. The number of rows is not part of the type and is not checked.

Tables from files. A table argument given as a file is read in the pool, not the nexus. Arrow IPC and Parquet are recognized by their magic bytes; otherwise the extension (.csv, .tsv, .json) or the content decides. CSV columns bind to declared columns by header name. As output, the nexus writes a table as JSON rows, JSON lines, CSV, TSV, Arrow IPC or Parquet.

Tensors are not a runtime type. A tensor is a pair of its dimensions and its elements in one flat, row-major array, and travels as that ordinary voidstar tuple: Matrix 2 2 Real has schema t2t2jja:4f8. Because the array of numbers is 64-byte aligned in shared memory, Python receives a read-only numpy view over it with no copy. When the dimensions are known at compile time, the array’s length is checked against their product.