static tdb_error store_lexicons(tdb_cons *cons)
{
tdb_field i;
FILE *out = NULL;
char path[TDB_MAX_PATH_SIZE];
int ret = 0;
TDB_PATH(path, "%s/fields", cons->root);
TDB_OPEN(out, path, "w");
for (i = 0; i < cons->num_ofields; i++){
TDB_PATH(path, "%s/lexicon.%s", cons->root, cons->ofield_names[i]);
if ((ret = lexicon_store(&cons->lexicons[i], path)))
goto done;
TDB_FPRINTF(out, "%s\n", cons->ofield_names[i]);
}
TDB_FPRINTF(out, "\n");
done:
TDB_CLOSE_FINAL(out);
return ret;
}
static tdb_error store_version(tdb_cons *cons)
{
FILE *out = NULL;
char path[TDB_MAX_PATH_SIZE];
int ret = 0;
TDB_PATH(path, "%s/version", cons->root);
TDB_OPEN(out, path, "w");
TDB_FPRINTF(out, "%llu", TDB_VERSION_LATEST);
done:
TDB_CLOSE_FINAL(out);
return ret;
}
static tdb_error store_uuids(tdb_cons *cons)
{
char path[TDB_MAX_PATH_SIZE];
struct jm_fold_state state = {.ret = 0};
uint64_t num_trails = j128m_num_keys(&cons->trails);
int ret = 0;
/* this is why num_trails < TDB_MAX)NUM_TRAILS < 2^59:
(2^59 - 1) * 16 < LONG_MAX (off_t) */
if (num_trails > TDB_MAX_NUM_TRAILS)
return TDB_ERR_TOO_MANY_TRAILS;
TDB_PATH(path, "%s/uuids", cons->root);
TDB_OPEN(state.out, path, "w");
TDB_TRUNCATE(state.out, ((off_t)(num_trails * 16)));
j128m_fold(&cons->trails, store_uuids_fun, &state);
ret = state.ret;
done:
TDB_CLOSE_FINAL(state.out);
return ret;
}
int is_fieldname_invalid(const char* field)
{
uint64_t i;
if (!strcmp(field, "time"))
return 1;
for (i = 0; i < TDB_MAX_FIELDNAME_LENGTH && field[i]; i++)
if (!index(TDB_FIELDNAME_CHARS, field[i]))
return 1;
if (i == 0 || i == TDB_MAX_FIELDNAME_LENGTH)
return 1;
return 0;
}
TDB_EXPORT tdb_error tdb_cons_open(tdb_cons *cons,
const char *root,
const char **ofield_names,
uint64_t num_ofields)
{
tdb_field i;
int fd;
int ret = 0;
/*
by handling the "cons == NULL" case here gracefully, we allow the return
value of tdb_init() to be used unchecked like here:
int err;
tdb_cons *cons = tdb_cons_init();
if ((err = tdb_cons_open(cons, path, fields, num_fields)))
printf("Opening cons failed: %s", tdb_error(err));
*/
if (!cons)
return TDB_ERR_HANDLE_IS_NULL;
if (cons->events.item_size)
return TDB_ERR_HANDLE_ALREADY_OPENED;
if (num_ofields > TDB_MAX_NUM_FIELDS)
return TDB_ERR_TOO_MANY_FIELDS;
if ((ret = find_duplicate_fieldnames(ofield_names, num_ofields)))
goto done;
if (!(cons->ofield_names = calloc(num_ofields, sizeof(char*))))
return TDB_ERR_NOMEM;
for (i = 0; i < num_ofields; i++){
if (is_fieldname_invalid(ofield_names[i])){
ret = TDB_ERR_INVALID_FIELDNAME;
goto done;
}
if (!(cons->ofield_names[i] = strdup(ofield_names[i]))){
ret = TDB_ERR_NOMEM;
goto done;
}
}
j128m_init(&cons->trails);
if (!(cons->root = strdup(root))){
ret = TDB_ERR_NOMEM;
goto done;
}
cons->min_timestamp = UINT64_MAX;
cons->num_ofields = num_ofields;
cons->events.arena_increment = EVENTS_ARENA_INCREMENT;
cons->events.item_size = sizeof(struct tdb_cons_event);
cons->items.item_size = sizeof(tdb_item);
/* Opportunistically try to create the output directory.
We don't care if it fails, e.g. because it already exists */
mkdir(root, 0755);
TDB_PATH(cons->tempfile, "%s/tmp.items.XXXXXX", root);
if ((fd = mkstemp(cons->tempfile)) == -1){
ret = TDB_ERR_IO_OPEN;
goto done;
}
if (!(cons->items.fd = fdopen(fd, "w"))){
ret = TDB_ERR_IO_OPEN;
goto done;
}
if (cons->num_ofields > 0)
if (!(cons->lexicons = calloc(cons->num_ofields,
sizeof(struct judy_str_map)))){
ret = TDB_ERR_NOMEM;
goto done;
}
for (i = 0; i < cons->num_ofields; i++)
if (jsm_init(&cons->lexicons[i])){
ret = TDB_ERR_NOMEM;
goto done;
}
done:
return ret;
}
tdb_error tdb_encode(tdb_cons *cons, const tdb_item *items)
{
char path[TDB_MAX_PATH_SIZE];
char grouped_path[TDB_MAX_PATH_SIZE];
char toc_path[TDB_MAX_PATH_SIZE];
char *root = cons->root;
char *read_buf = NULL;
struct field_stats *fstats = NULL;
uint64_t num_trails = 0;
uint64_t num_events = cons->events.next;
uint64_t num_fields = cons->num_ofields + 1;
uint64_t max_timestamp = 0;
uint64_t max_timedelta = 0;
uint64_t *field_cardinalities = NULL;
uint64_t i;
Pvoid_t unigram_freqs = NULL;
struct judy_128_map gram_freqs;
struct judy_128_map codemap;
Word_t tmp;
FILE *grouped_w = NULL;
FILE *grouped_r = NULL;
int fd, ret = 0;
TDB_TIMER_DEF
j128m_init(&gram_freqs);
j128m_init(&codemap);
if (!(field_cardinalities = calloc(cons->num_ofields, 8))){
ret = TDB_ERR_NOMEM;
goto done;
}
for (i = 0; i < cons->num_ofields; i++)
field_cardinalities[i] = jsm_num_keys(&cons->lexicons[i]);
/* 1. group events by trail, sort events of each trail by time,
and delta-encode timestamps */
TDB_TIMER_START
TDB_PATH(grouped_path, "%s/tmp.grouped.XXXXXX", root);
if ((fd = mkstemp(grouped_path)) == -1){
ret = TDB_ERR_IO_OPEN;
goto done;
}
if (!(grouped_w = fdopen(fd, "w"))){
ret = TDB_ERR_IO_OPEN;
goto done;
}
if (cons->events.data)
if ((ret = groupby_uuid(grouped_w,
(struct tdb_cons_event*)cons->events.data,
cons,
&num_trails,
&max_timestamp,
&max_timedelta)))
goto done;
/*
not the most clean separation of ownership here, but these objects
can be huge so keeping them around unecessarily is expensive
*/
free(cons->events.data);
cons->events.data = NULL;
j128m_free(&cons->trails);
TDB_CLOSE(grouped_w);
grouped_w = NULL;
TDB_OPEN(grouped_r, grouped_path, "r");
if (!(read_buf = malloc(READ_BUFFER_SIZE))){
ret = TDB_ERR_NOMEM;
goto done;
}
setvbuf(grouped_r, read_buf, _IOFBF, READ_BUFFER_SIZE);
TDB_TIMER_END("trail/groupby_uuid");
/* 2. store metatadata */
TDB_TIMER_START
TDB_PATH(path, "%s/info", root);
if ((ret = store_info(path,
num_trails,
num_events,
cons->min_timestamp,
max_timestamp,
max_timedelta)))
goto done;
TDB_TIMER_END("trail/info");
/* 3. collect value (unigram) freqs, including delta-encoded timestamps */
TDB_TIMER_START
unigram_freqs = collect_unigrams(grouped_r, num_events, items, num_fields);
if (num_events > 0 && !unigram_freqs){
ret = TDB_ERR_NOMEM;
goto done;
}
TDB_TIMER_END("trail/collect_unigrams");
/* 4. construct uni/bi-grams */
tdb_opt_value dont_build_bigrams;
tdb_cons_get_opt(cons, TDB_OPT_CONS_NO_BIGRAMS, &dont_build_bigrams);
TDB_TIMER_START
if ((ret = make_grams(grouped_r,
num_events,
items,
num_fields,
unigram_freqs,
&gram_freqs,
dont_build_bigrams.value)))
goto done;
TDB_TIMER_END("trail/gram_freqs");
/* 5. build a huffman codebook and stats struct for encoding grams */
TDB_TIMER_START
if ((ret = huff_create_codemap(&gram_freqs, &codemap)))
goto done;
if (!(fstats = huff_field_stats(field_cardinalities,
num_fields,
max_timedelta))){
ret = TDB_ERR_NOMEM;
goto done;
}
TDB_TIMER_END("trail/huff_create_codemap");
/* 6. encode and write trails to disk */
TDB_TIMER_START
TDB_PATH(path, "%s/trails.data", root);
TDB_PATH(toc_path, "%s/trails.toc", root);
if ((ret = encode_trails(items,
grouped_r,
num_events,
num_trails,
num_fields,
&codemap,
&gram_freqs,
fstats,
path,
toc_path)))
goto done;
TDB_TIMER_END("trail/encode_trails");
/* 7. write huffman codebook to disk */
TDB_TIMER_START
tdb_path(path, "%s/trails.codebook", root);
if ((ret = store_codebook(&codemap, path)))
goto done;
TDB_TIMER_END("trail/store_codebook");
done:
TDB_CLOSE_FINAL(grouped_w);
TDB_CLOSE_FINAL(grouped_r);
j128m_free(&gram_freqs);
j128m_free(&codemap);
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wsign-compare"
JLFA(tmp, unigram_freqs);
#pragma GCC diagnostic pop
unlink(grouped_path);
free(field_cardinalities);
free(read_buf);
free(fstats);
return ret;
out_of_memory:
return TDB_ERR_NOMEM;
}
