BUG: Don't over-optimize memory with jagged CSV

With jagged CSV's, we risk being too quick
to dump memory that we need to allocate
because previous chunks would have
indicated much larger rows than we can
anticipate in subsequent chunks.

Closes gh-23509.

Author: gfyoung

doc/source/whatsnew/v0.24.0.txt

@@ -1298,6 +1298,7 @@ Notice how we now instead output ``np.nan`` itself instead of a stringified form
 - :func:`read_excel()` will correctly show the deprecation warning for previously deprecated ``sheetname`` (:issue:`17994`)
 - :func:`read_csv()` and func:`read_table()` will throw ``UnicodeError`` and not coredump on badly encoded strings (:issue:`22748`)
 - :func:`read_csv()` will correctly parse timezone-aware datetimes (:issue:`22256`)
+- Bug in :func:`read_csv()` in which memory management was prematurely optimized for the C engine when the data was being read in chunks (:issue:`23509`)
 - :func:`read_sas()` will parse numbers in sas7bdat-files that have width less than 8 bytes correctly. (:issue:`21616`)
 - :func:`read_sas()` will correctly parse sas7bdat files with many columns (:issue:`22628`)
 - :func:`read_sas()` will correctly parse sas7bdat files with data page types having also bit 7 set (so page type is 128 + 256 = 384) (:issue:`16615`)

pandas/_libs/parsers.pyx

@@ -132,6 +132,7 @@ cdef extern from "parser/tokenizer.h":
         int64_t *word_starts  # where we are in the stream
         int64_t words_len
         int64_t words_cap
+        int64_t max_words_cap    # maximum word cap encountered
 
         char *pword_start        # pointer to stream start of current field
         int64_t word_start       # position start of current field

pandas/_libs/src/parser/tokenizer.c

@@ -197,6 +197,7 @@ int parser_init(parser_t *self) {
     sz = sz ? sz : 1;
     self->words = (char **)malloc(sz * sizeof(char *));
     self->word_starts = (int64_t *)malloc(sz * sizeof(int64_t));
+    self->max_words_cap = sz;
     self->words_cap = sz;
     self->words_len = 0;
 
@@ -247,7 +248,7 @@ void parser_del(parser_t *self) {
 }
 
 static int make_stream_space(parser_t *self, size_t nbytes) {
-    int64_t i, cap;
+    int64_t i, cap, length;
     int status;
     void *orig_ptr, *newptr;
 
@@ -287,8 +288,23 @@ static int make_stream_space(parser_t *self, size_t nbytes) {
     */
 
     cap = self->words_cap;
+
+    /**
+     * If we are reading in chunks, we need to be aware of the maximum number
+     * of words we have seen in previous chunks (self->max_words_cap), so
+     * that way, we can properly allocate when reading subsequent ones.
+     *
+     * Otherwise, we risk a buffer overflow if we mistakenly under-allocate
+     * just because a recent chunk did not have as many words.
+     */
+    if (self->words_len + nbytes < self->max_words_cap) {
+        length = self->max_words_cap - nbytes;
+    } else {
+        length = self->words_len;
+    }
+
     self->words =
-        (char **)grow_buffer((void *)self->words, self->words_len,
+        (char **)grow_buffer((void *)self->words, length,
                              (int64_t*)&self->words_cap, nbytes,
                              sizeof(char *), &status);
     TRACE(
@@ -1241,6 +1257,19 @@ int parser_trim_buffers(parser_t *self) {
 
     int64_t i;
 
+    /**
+     * Before we free up space and trim, we should
+     * save how many words we saw when parsing, if
+     * it exceeds the maximum number we saw before.
+     *
+     * This is important for when we read in chunks,
+     * so that we can inform subsequent chunk parsing
+     * as to how many words we could possibly see.
+     */
+    if (self->words_cap > self->max_words_cap) {
+        self->max_words_cap = self->words_cap;
+    }
+
     /* trim words, word_starts */
     new_cap = _next_pow2(self->words_len) + 1;
     if (new_cap < self->words_cap) {

pandas/_libs/src/parser/tokenizer.h

@@ -142,6 +142,7 @@ typedef struct parser_t {
     int64_t *word_starts;   // where we are in the stream
     int64_t words_len;
     int64_t words_cap;
+    int64_t max_words_cap;  // maximum word cap encountered
 
     char *pword_start;      // pointer to stream start of current field
     int64_t word_start;     // position start of current field

pandas/tests/io/parser/common.py

@@ -458,6 +458,22 @@ def test_read_chunksize_generated_index(self):
 
         tm.assert_frame_equal(pd.concat(reader), df)
 
+    def test_read_chunksize_jagged_names(self):
+        # see gh-23509
+        data = "\n".join(["0"] * 7 + [",".join(["0"] * 10)])
+        reader = self.read_csv(StringIO(data), names=range(10), chunksize=4)
+
+        expected = DataFrame()
+
+        for i in range(10):
+            if i == 0:
+                expected[i] = [0] * 8
+            else:
+                expected[i] = [np.nan] * 7 + [0]
+
+        result = pd.concat(reader)
+        tm.assert_frame_equal(result, expected)
+
     def test_read_text_list(self):
         data = """A,B,C\nfoo,1,2,3\nbar,4,5,6"""
         as_list = [['A', 'B', 'C'], ['foo', '1', '2', '3'], ['bar',