tranfi 0.0.2 → 0.1.2

package/csrc/op_stats.c

@@ -0,0 +1,779 @@
+/*
+ * op_stats.c — Streaming aggregates with online algorithms.
+ *
+ * Config: {"stats": ["count","sum","avg","min","max","var","stddev",
+ *                     "median","p25","p75","skewness","kurtosis",
+ *                     "distinct","hist","sample"]}
+ *   or {} for default stats (count, sum, avg, min, max, var, stddev, median).
+ *
+ * Algorithms ported from OnlineStats.jl (MIT):
+ *   - Welford's online variance (var, stddev)
+ *   - Moments via non-central moment tracking (skewness, kurtosis)
+ *   - P2 quantile — Jain-Chlamtac (median, p25, p75)
+ *   - HyperLogLog cardinality estimation (distinct)
+ *   - Streaming histogram with adaptive bins (hist)
+ *   - Reservoir sampling — Algorithm R (sample)
+ *
+ * Output: long format, one row per input column.
+ */
+
+#include "internal.h"
+#include "cJSON.h"
+#include "date_utils.h"
+#include <stdlib.h>
+#include <string.h>
+#include <stdio.h>
+#include <float.h>
+#include <math.h>
+
+/* ---- P2 Quantile (Jain-Chlamtac) ---- */
+
+typedef struct {
+    double q[5];       /* marker heights */
+    int    n[5];       /* marker positions */
+    double np[5];      /* desired positions */
+    double dn[5];      /* increment for desired positions */
+    double tau;        /* target quantile */
+    int    nobs;
+    double init[5];    /* first 5 observations for init */
+} p2_quantile;
+
+static void p2_init(p2_quantile *p, double tau) {
+    memset(p, 0, sizeof(*p));
+    p->tau = tau;
+    p->dn[0] = 0;
+    p->dn[1] = tau / 2.0;
+    p->dn[2] = tau;
+    p->dn[3] = (1.0 + tau) / 2.0;
+    p->dn[4] = 1.0;
+}
+
+static int dbl_cmp(const void *a, const void *b) {
+    double da = *(const double *)a, db = *(const double *)b;
+    return (da > db) - (da < db);
+}
+
+static void p2_interpolate(double *q, int *n, int i, int d) {
+    double df = (double)d;
+    double qi = q[i-1], qm = q[i], qp = q[i+1];
+    int ni = n[i-1], nm = n[i], np = n[i+1];
+
+    double v1 = (double)(nm - ni + d) * (qp - qm) / (double)(np - nm);
+    double v2 = (double)(np - nm - d) * (qm - qi) / (double)(nm - ni);
+    double qnew = qm + df / (double)(np - ni) * (v1 + v2);
+
+    if (qi < qnew && qnew < qp) {
+        q[i] = qnew;
+    } else {
+        /* linear fallback */
+        q[i] = qm + df * (q[i + d] - qm) / (double)(n[i + d] - nm);
+    }
+    n[i] += d;
+}
+
+static void p2_update(p2_quantile *p, double y) {
+    p->nobs++;
+    if (p->nobs <= 5) {
+        p->init[p->nobs - 1] = y;
+        if (p->nobs == 5) {
+            qsort(p->init, 5, sizeof(double), dbl_cmp);
+            for (int i = 0; i < 5; i++) {
+                p->q[i] = p->init[i];
+                p->n[i] = i + 1;
+            }
+            p->np[0] = 1;
+            p->np[1] = 1 + 2 * p->tau;
+            p->np[2] = 1 + 4 * p->tau;
+            p->np[3] = 3 + 2 * p->tau;
+            p->np[4] = 5;
+        }
+        return;
+    }
+
+    /* Find cell k */
+    int k;
+    if (y < p->q[0]) {
+        p->q[0] = y;
+        k = 0;
+    } else if (y >= p->q[4]) {
+        p->q[4] = y > p->q[4] ? y : p->q[4];
+        k = 3;
+    } else {
+        k = 0;
+        for (int i = 1; i < 5; i++) {
+            if (y < p->q[i]) { k = i - 1; break; }
+        }
+    }
+
+    /* Increment positions */
+    for (int i = k + 1; i < 5; i++) p->n[i]++;
+    for (int i = 0; i < 5; i++) p->np[i] += p->dn[i];
+
+    /* Adjust markers 1-3 */
+    for (int i = 1; i <= 3; i++) {
+        double di = p->np[i] - (double)p->n[i];
+        if ((di >= 1.0 && p->n[i+1] - p->n[i] > 1) ||
+            (di <= -1.0 && p->n[i-1] - p->n[i] < -1)) {
+            int d = di >= 0 ? 1 : -1;
+            p2_interpolate(p->q, p->n, i, d);
+        }
+    }
+}
+
+static double p2_value(const p2_quantile *p) {
+    if (p->nobs < 5) {
+        /* Not enough data; sort what we have and pick quantile */
+        if (p->nobs == 0) return NAN;
+        double tmp[5];
+        memcpy(tmp, p->init, p->nobs * sizeof(double));
+        qsort(tmp, p->nobs, sizeof(double), dbl_cmp);
+        size_t idx = (size_t)(p->tau * (p->nobs - 1));
+        return tmp[idx];
+    }
+    return p->q[2]; /* middle marker */
+}
+
+/* ---- HyperLogLog ---- */
+
+#define HLL_P 10
+#define HLL_M (1 << HLL_P)  /* 1024 registers */
+
+typedef struct {
+    uint8_t M[HLL_M];
+    size_t  n;
+} hll_state;
+
+static void hll_init(hll_state *h) {
+    memset(h, 0, sizeof(*h));
+}
+
+static uint32_t hll_hash(const char *key) {
+    /* MurmurHash3 finalizer-style mixing */
+    uint32_t h = 0x811c9dc5u;
+    for (const char *p = key; *p; p++) {
+        h ^= (uint32_t)(unsigned char)*p;
+        h *= 0x01000193u;
+    }
+    h ^= h >> 16;
+    h *= 0x85ebca6bu;
+    h ^= h >> 13;
+    h *= 0xc2b2ae35u;
+    h ^= h >> 16;
+    return h;
+}
+
+static int hll_clz32(uint32_t x) {
+    if (x == 0) return 32;
+#if defined(__GNUC__) || defined(__clang__)
+    return __builtin_clz(x);
+#else
+    int n = 0;
+    if (!(x & 0xFFFF0000u)) { n += 16; x <<= 16; }
+    if (!(x & 0xFF000000u)) { n +=  8; x <<=  8; }
+    if (!(x & 0xF0000000u)) { n +=  4; x <<=  4; }
+    if (!(x & 0xC0000000u)) { n +=  2; x <<=  2; }
+    if (!(x & 0x80000000u)) { n +=  1; }
+    return n;
+#endif
+}
+
+static void hll_update(hll_state *h, const char *val) {
+    h->n++;
+    uint32_t x = hll_hash(val);
+    uint32_t idx = x & (HLL_M - 1);
+    uint32_t w = x >> HLL_P;
+    int rho = hll_clz32(w) + 1;
+    if (rho > 32 - HLL_P) rho = 32 - HLL_P;
+    if ((uint8_t)rho > h->M[idx]) h->M[idx] = (uint8_t)rho;
+}
+
+static double hll_estimate(const hll_state *h) {
+    double alpha;
+    if (HLL_P == 4)      alpha = 0.673;
+    else if (HLL_P == 5) alpha = 0.697;
+    else if (HLL_P == 6) alpha = 0.709;
+    else                  alpha = 0.7213 / (1.0 + 1.079 / HLL_M);
+
+    double sum = 0;
+    int V = 0;
+    for (int i = 0; i < HLL_M; i++) {
+        sum += 1.0 / (1u << h->M[i]);
+        if (h->M[i] == 0) V++;
+    }
+    double E = alpha * HLL_M * HLL_M / sum;
+
+    /* Small/large range corrections */
+    if (E <= 2.5 * HLL_M) {
+        if (V > 0) return HLL_M * log((double)HLL_M / V);
+        return E;
+    }
+    if (E <= (1.0 / 30.0) * 4294967296.0) return E;
+    return -4294967296.0 * log(1.0 - E / 4294967296.0);
+}
+
+/* ---- Streaming Histogram (adaptive bins) ---- */
+
+#define HIST_NBINS 32
+
+typedef struct {
+    double edges[HIST_NBINS + 1];
+    size_t counts[HIST_NBINS];
+    size_t n;
+    double lo, hi;
+    int    initialized;
+} hist_state;
+
+static void hist_init(hist_state *h) {
+    memset(h, 0, sizeof(*h));
+}
+
+static void hist_setup_edges(hist_state *h) {
+    double step = (h->hi - h->lo) / HIST_NBINS;
+    for (int i = 0; i <= HIST_NBINS; i++)
+        h->edges[i] = h->lo + step * i;
+}
+
+static int hist_bin(const hist_state *h, double y) {
+    if (y < h->edges[0]) return -1;
+    if (y >= h->edges[HIST_NBINS]) return HIST_NBINS - 1; /* last bin is closed */
+    /* binary search */
+    int lo = 0, hi = HIST_NBINS - 1;
+    while (lo < hi) {
+        int mid = (lo + hi) / 2;
+        if (y < h->edges[mid + 1]) hi = mid;
+        else lo = mid + 1;
+    }
+    return lo;
+}
+
+static void hist_expand(hist_state *h, double y) {
+    if (y >= h->lo && y <= h->hi) return;
+
+    if (y > h->hi) {
+        /* Expand right: double range until y fits */
+        while (y > h->hi) {
+            double w = h->hi - h->lo;
+            h->hi = h->lo + w * 2.0;
+            /* Merge bin pairs: bins [0,1]→0, [2,3]→1, etc. */
+            for (int i = 0; i < HIST_NBINS / 2; i++)
+                h->counts[i] = h->counts[2*i] + h->counts[2*i + 1];
+            for (int i = HIST_NBINS / 2; i < HIST_NBINS; i++)
+                h->counts[i] = 0;
+            hist_setup_edges(h);
+        }
+    } else {
+        /* Expand left */
+        while (y < h->lo) {
+            double w = h->hi - h->lo;
+            h->lo = h->hi - w * 2.0;
+            for (int i = HIST_NBINS - 1; i >= HIST_NBINS / 2; i--)
+                h->counts[i] = h->counts[2*(i - HIST_NBINS/2)] +
+                                h->counts[2*(i - HIST_NBINS/2) + 1];
+            for (int i = 0; i < HIST_NBINS / 2; i++)
+                h->counts[i] = 0;
+            hist_setup_edges(h);
+        }
+    }
+}
+
+static void hist_update(hist_state *h, double y) {
+    h->n++;
+    if (!h->initialized) {
+        h->lo = y;
+        h->hi = y;
+        h->initialized = 1;
+        return; /* single value, no bin yet */
+    }
+    if (h->lo == h->hi) {
+        /* Second distinct value */
+        if (y == h->lo) return;
+        if (y < h->lo) h->lo = y; else h->hi = y;
+        /* Put prior observations in appropriate bin */
+        double range = h->hi - h->lo;
+        h->lo -= range * 0.01;
+        h->hi += range * 0.01;
+        hist_setup_edges(h);
+        /* All prior n-1 points were the same value, assign to their bin */
+        int b = hist_bin(h, h->lo + range * 0.01 + range * 0.5);
+        if (b >= 0 && b < HIST_NBINS) h->counts[b] = h->n - 1;
+        b = hist_bin(h, y);
+        if (b >= 0 && b < HIST_NBINS) h->counts[b]++;
+        return;
+    }
+
+    hist_expand(h, y);
+    int b = hist_bin(h, y);
+    if (b >= 0 && b < HIST_NBINS) h->counts[b]++;
+}
+
+/* ---- Reservoir Sample (Algorithm R) ---- */
+
+#define RESERVOIR_K 10
+
+typedef struct {
+    double values[RESERVOIR_K];
+    size_t n;
+    /* Simple xorshift RNG state */
+    uint64_t rng;
+} reservoir_state;
+
+static void reservoir_init(reservoir_state *r) {
+    memset(r, 0, sizeof(*r));
+    r->rng = 0x12345678deadbeefULL;
+}
+
+static uint64_t reservoir_rand(reservoir_state *r) {
+    r->rng ^= r->rng << 13;
+    r->rng ^= r->rng >> 7;
+    r->rng ^= r->rng << 17;
+    return r->rng;
+}
+
+static void reservoir_update(reservoir_state *r, double y) {
+    r->n++;
+    if (r->n <= RESERVOIR_K) {
+        r->values[r->n - 1] = y;
+    } else {
+        uint64_t j = reservoir_rand(r) % r->n;
+        if (j < RESERVOIR_K) {
+            r->values[j] = y;
+        }
+    }
+}
+
+/* ---- Per-column accumulator ---- */
+
+typedef struct {
+    /* Basic stats */
+    size_t count;
+    double sum;
+    double min;
+    double max;
+    int    has_numeric;
+
+    /* Welford's variance */
+    double wf_mean;
+    double wf_m2;      /* sum of (x - mean)^2 */
+
+    /* Moments (non-central) for skewness/kurtosis */
+    double mom[4];      /* E[X], E[X^2], E[X^3], E[X^4] */
+
+    /* P2 quantiles */
+    p2_quantile p2_median;
+    p2_quantile p2_p25;
+    p2_quantile p2_p75;
+
+    /* HyperLogLog */
+    hll_state *hll;
+
+    /* Histogram */
+    hist_state *hist;
+
+    /* Reservoir sample */
+    reservoir_state *reservoir;
+} col_accum;
+
+/* ---- Stats state ---- */
+
+typedef struct {
+    col_accum  *accums;
+    char      **col_names;
+    size_t      n_cols;
+    int         initialized;
+    /* Which stats to output */
+    int want_count, want_sum, want_avg, want_min, want_max;
+    int want_var, want_stddev;
+    int want_median, want_p25, want_p75;
+    int want_skewness, want_kurtosis;
+    int want_distinct;
+    int want_hist;
+    int want_sample;
+} stats_state;
+
+static void accum_init(col_accum *a, const stats_state *st) {
+    a->min = DBL_MAX;
+    a->max = -DBL_MAX;
+    p2_init(&a->p2_median, 0.5);
+    p2_init(&a->p2_p25, 0.25);
+    p2_init(&a->p2_p75, 0.75);
+
+    if (st->want_distinct) {
+        a->hll = calloc(1, sizeof(hll_state));
+        if (a->hll) hll_init(a->hll);
+    }
+    if (st->want_hist) {
+        a->hist = calloc(1, sizeof(hist_state));
+        if (a->hist) hist_init(a->hist);
+    }
+    if (st->want_sample) {
+        a->reservoir = calloc(1, sizeof(reservoir_state));
+        if (a->reservoir) reservoir_init(a->reservoir);
+    }
+}
+
+static void accum_update(col_accum *a, double val, const char *str_val,
+                          int is_num) {
+    a->count++;
+    if (is_num) {
+        a->has_numeric = 1;
+        a->sum += val;
+        if (val < a->min) a->min = val;
+        if (val > a->max) a->max = val;
+
+        /* Welford */
+        double old_mean = a->wf_mean;
+        a->wf_mean += (val - a->wf_mean) / (double)a->count;
+        a->wf_m2 += (val - a->wf_mean) * (val - old_mean);
+
+        /* Moments */
+        double gamma = 1.0 / (double)a->count;
+        double y2 = val * val;
+        a->mom[0] += gamma * (val      - a->mom[0]);
+        a->mom[1] += gamma * (y2       - a->mom[1]);
+        a->mom[2] += gamma * (val * y2 - a->mom[2]);
+        a->mom[3] += gamma * (y2 * y2  - a->mom[3]);
+
+        /* P2 quantiles */
+        p2_update(&a->p2_median, val);
+        p2_update(&a->p2_p25, val);
+        p2_update(&a->p2_p75, val);
+
+        /* Histogram */
+        if (a->hist) hist_update(a->hist, val);
+
+        /* Reservoir */
+        if (a->reservoir) reservoir_update(a->reservoir, val);
+    }
+
+    /* HyperLogLog uses string representation for all types */
+    if (a->hll && str_val) hll_update(a->hll, str_val);
+}
+
+static void accum_free(col_accum *a) {
+    free(a->hll);
+    free(a->hist);
+    free(a->reservoir);
+}
+
+/* ---- Process / Flush / Destroy ---- */
+
+static int stats_process(tf_step *self, tf_batch *in, tf_batch **out,
+                         tf_side_channels *side) {
+    (void)side;
+    stats_state *st = self->state;
+    *out = NULL;
+
+    if (!st->initialized) {
+        st->n_cols = in->n_cols;
+        st->accums = calloc(in->n_cols, sizeof(col_accum));
+        st->col_names = calloc(in->n_cols, sizeof(char *));
+        if (!st->accums || !st->col_names) return TF_ERROR;
+        for (size_t c = 0; c < in->n_cols; c++) {
+            st->col_names[c] = strdup(in->col_names[c]);
+            accum_init(&st->accums[c], st);
+        }
+        st->initialized = 1;
+    }
+
+    for (size_t r = 0; r < in->n_rows; r++) {
+        for (size_t c = 0; c < st->n_cols && c < in->n_cols; c++) {
+            if (tf_batch_is_null(in, r, c)) continue;
+
+            double val = 0;
+            int is_num = 0;
+            char str_buf[64];
+            const char *str_val = NULL;
+
+            switch (in->col_types[c]) {
+                case TF_TYPE_INT64: {
+                    int64_t iv = tf_batch_get_int64(in, r, c);
+                    val = (double)iv;
+                    is_num = 1;
+                    snprintf(str_buf, sizeof(str_buf), "%lld", (long long)iv);
+                    str_val = str_buf;
+                    break;
+                }
+                case TF_TYPE_FLOAT64:
+                    val = tf_batch_get_float64(in, r, c);
+                    is_num = 1;
+                    snprintf(str_buf, sizeof(str_buf), "%.17g", val);
+                    str_val = str_buf;
+                    break;
+                case TF_TYPE_STRING:
+                    str_val = tf_batch_get_string(in, r, c);
+                    break;
+                case TF_TYPE_BOOL:
+                    str_val = tf_batch_get_bool(in, r, c) ? "true" : "false";
+                    break;
+                case TF_TYPE_DATE: {
+                    int32_t dv = tf_batch_get_date(in, r, c);
+                    val = (double)dv;
+                    is_num = 1;
+                    snprintf(str_buf, sizeof(str_buf), "%d", (int)dv);
+                    str_val = str_buf;
+                    break;
+                }
+                case TF_TYPE_TIMESTAMP: {
+                    int64_t tv = tf_batch_get_timestamp(in, r, c);
+                    val = (double)tv;
+                    is_num = 1;
+                    snprintf(str_buf, sizeof(str_buf), "%lld", (long long)tv);
+                    str_val = str_buf;
+                    break;
+                }
+                default:
+                    break;
+            }
+            accum_update(&st->accums[c], val, str_val, is_num);
+        }
+    }
+
+    return TF_OK;
+}
+
+static int stats_flush(tf_step *self, tf_batch **out, tf_side_channels *side) {
+    (void)side;
+    stats_state *st = self->state;
+    *out = NULL;
+
+    if (!st->initialized || st->n_cols == 0) return TF_OK;
+
+    /* Count output columns */
+    size_t n_stat_cols = 1; /* "column" */
+    if (st->want_count)    n_stat_cols++;
+    if (st->want_sum)      n_stat_cols++;
+    if (st->want_avg)      n_stat_cols++;
+    if (st->want_min)      n_stat_cols++;
+    if (st->want_max)      n_stat_cols++;
+    if (st->want_var)      n_stat_cols++;
+    if (st->want_stddev)   n_stat_cols++;
+    if (st->want_median)   n_stat_cols++;
+    if (st->want_p25)      n_stat_cols++;
+    if (st->want_p75)      n_stat_cols++;
+    if (st->want_skewness) n_stat_cols++;
+    if (st->want_kurtosis) n_stat_cols++;
+    if (st->want_distinct) n_stat_cols++;
+    if (st->want_hist)     n_stat_cols++;
+    if (st->want_sample)   n_stat_cols++;
+
+    tf_batch *ob = tf_batch_create(n_stat_cols, st->n_cols);
+    if (!ob) return TF_ERROR;
+
+    /* Set schema */
+    size_t ci = 0;
+    tf_batch_set_schema(ob, ci++, "column", TF_TYPE_STRING);
+    if (st->want_count)    tf_batch_set_schema(ob, ci++, "count", TF_TYPE_INT64);
+    if (st->want_sum)      tf_batch_set_schema(ob, ci++, "sum", TF_TYPE_FLOAT64);
+    if (st->want_avg)      tf_batch_set_schema(ob, ci++, "avg", TF_TYPE_FLOAT64);
+    if (st->want_min)      tf_batch_set_schema(ob, ci++, "min", TF_TYPE_FLOAT64);
+    if (st->want_max)      tf_batch_set_schema(ob, ci++, "max", TF_TYPE_FLOAT64);
+    if (st->want_var)      tf_batch_set_schema(ob, ci++, "var", TF_TYPE_FLOAT64);
+    if (st->want_stddev)   tf_batch_set_schema(ob, ci++, "stddev", TF_TYPE_FLOAT64);
+    if (st->want_median)   tf_batch_set_schema(ob, ci++, "median", TF_TYPE_FLOAT64);
+    if (st->want_p25)      tf_batch_set_schema(ob, ci++, "p25", TF_TYPE_FLOAT64);
+    if (st->want_p75)      tf_batch_set_schema(ob, ci++, "p75", TF_TYPE_FLOAT64);
+    if (st->want_skewness) tf_batch_set_schema(ob, ci++, "skewness", TF_TYPE_FLOAT64);
+    if (st->want_kurtosis) tf_batch_set_schema(ob, ci++, "kurtosis", TF_TYPE_FLOAT64);
+    if (st->want_distinct) tf_batch_set_schema(ob, ci++, "distinct", TF_TYPE_INT64);
+    if (st->want_hist)     tf_batch_set_schema(ob, ci++, "hist", TF_TYPE_STRING);
+    if (st->want_sample)   tf_batch_set_schema(ob, ci++, "sample", TF_TYPE_STRING);
+
+    /* One row per input column */
+    for (size_t c = 0; c < st->n_cols; c++) {
+        tf_batch_ensure_capacity(ob, c + 1);
+        ci = 0;
+        tf_batch_set_string(ob, c, ci++, st->col_names[c]);
+
+        col_accum *a = &st->accums[c];
+
+        if (st->want_count) {
+            tf_batch_set_int64(ob, c, ci++, (int64_t)a->count);
+        }
+        if (st->want_sum) {
+            if (a->has_numeric) tf_batch_set_float64(ob, c, ci, a->sum);
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_avg) {
+            if (a->has_numeric && a->count > 0)
+                tf_batch_set_float64(ob, c, ci, a->sum / (double)a->count);
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_min) {
+            if (a->has_numeric && a->count > 0) tf_batch_set_float64(ob, c, ci, a->min);
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_max) {
+            if (a->has_numeric && a->count > 0) tf_batch_set_float64(ob, c, ci, a->max);
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_var) {
+            if (a->has_numeric && a->count > 1)
+                tf_batch_set_float64(ob, c, ci, a->wf_m2 / (double)(a->count - 1));
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_stddev) {
+            if (a->has_numeric && a->count > 1)
+                tf_batch_set_float64(ob, c, ci, sqrt(a->wf_m2 / (double)(a->count - 1)));
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_median) {
+            if (a->has_numeric && a->count > 0)
+                tf_batch_set_float64(ob, c, ci, p2_value(&a->p2_median));
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_p25) {
+            if (a->has_numeric && a->count > 0)
+                tf_batch_set_float64(ob, c, ci, p2_value(&a->p2_p25));
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_p75) {
+            if (a->has_numeric && a->count > 0)
+                tf_batch_set_float64(ob, c, ci, p2_value(&a->p2_p75));
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_skewness) {
+            if (a->has_numeric && a->count > 2) {
+                double m1 = a->mom[0], m2 = a->mom[1], m3 = a->mom[2];
+                double vr = m2 - m1 * m1;
+                if (vr > 1e-15) {
+                    double sk = (m3 - 3.0*m1*vr - m1*m1*m1) / pow(vr, 1.5);
+                    tf_batch_set_float64(ob, c, ci, sk);
+                } else {
+                    tf_batch_set_float64(ob, c, ci, 0.0);
+                }
+            } else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_kurtosis) {
+            if (a->has_numeric && a->count > 3) {
+                double m1 = a->mom[0], m2 = a->mom[1];
+                double m3 = a->mom[2], m4 = a->mom[3];
+                double vr = m2 - m1 * m1;
+                if (vr > 1e-15) {
+                    double kt = (m4 - 4.0*m1*m3 + 6.0*m1*m1*m2
+                                 - 3.0*m1*m1*m1*m1) / (vr*vr) - 3.0;
+                    tf_batch_set_float64(ob, c, ci, kt);
+                } else {
+                    tf_batch_set_float64(ob, c, ci, 0.0);
+                }
+            } else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_distinct) {
+            if (a->hll && a->count > 0)
+                tf_batch_set_int64(ob, c, ci, (int64_t)round(hll_estimate(a->hll)));
+            else tf_batch_set_null(ob, c, ci);
+            ci++;
+        }
+        if (st->want_hist) {
+            if (a->hist && a->hist->n > 1) {
+                /* Format: "lo:hi:c1,c2,...,cN" */
+                char buf[1024];
+                int pos = snprintf(buf, sizeof(buf), "%.6g:%.6g:",
+                                   a->hist->edges[0], a->hist->edges[HIST_NBINS]);
+                for (int b = 0; b < HIST_NBINS && pos < (int)sizeof(buf) - 16; b++) {
+                    if (b > 0) buf[pos++] = ',';
+                    pos += snprintf(buf + pos, sizeof(buf) - pos, "%zu", a->hist->counts[b]);
+                }
+                tf_batch_set_string(ob, c, ci, buf);
+            } else {
+                tf_batch_set_null(ob, c, ci);
+            }
+            ci++;
+        }
+        if (st->want_sample) {
+            if (a->reservoir && a->reservoir->n > 0) {
+                char buf[512];
+                int pos = 0;
+                size_t k = a->reservoir->n < RESERVOIR_K ? a->reservoir->n : RESERVOIR_K;
+                for (size_t s = 0; s < k && pos < (int)sizeof(buf) - 32; s++) {
+                    if (s > 0) buf[pos++] = ',';
+                    pos += snprintf(buf + pos, sizeof(buf) - pos,
+                                    "%.6g", a->reservoir->values[s]);
+                }
+                tf_batch_set_string(ob, c, ci, buf);
+            } else {
+                tf_batch_set_null(ob, c, ci);
+            }
+            ci++;
+        }
+        ob->n_rows = c + 1;
+    }
+
+    *out = ob;
+    return TF_OK;
+}
+
+static void stats_destroy(tf_step *self) {
+    stats_state *st = self->state;
+    if (st) {
+        if (st->accums) {
+            for (size_t i = 0; i < st->n_cols; i++)
+                accum_free(&st->accums[i]);
+            free(st->accums);
+        }
+        if (st->col_names) {
+            for (size_t i = 0; i < st->n_cols; i++) free(st->col_names[i]);
+            free(st->col_names);
+        }
+        free(st);
+    }
+    free(self);
+}
+
+tf_step *tf_stats_create(const cJSON *args) {
+    stats_state *st = calloc(1, sizeof(stats_state));
+    if (!st) return NULL;
+
+    cJSON *stats_arr = args ? cJSON_GetObjectItemCaseSensitive(args, "stats") : NULL;
+    if (stats_arr && cJSON_IsArray(stats_arr)) {
+        int n = cJSON_GetArraySize(stats_arr);
+        for (int i = 0; i < n; i++) {
+            cJSON *item = cJSON_GetArrayItem(stats_arr, i);
+            if (!cJSON_IsString(item)) continue;
+            const char *s = item->valuestring;
+            if (strcmp(s, "count") == 0)         st->want_count = 1;
+            else if (strcmp(s, "sum") == 0)      st->want_sum = 1;
+            else if (strcmp(s, "avg") == 0)      st->want_avg = 1;
+            else if (strcmp(s, "min") == 0)      st->want_min = 1;
+            else if (strcmp(s, "max") == 0)      st->want_max = 1;
+            else if (strcmp(s, "var") == 0)      st->want_var = 1;
+            else if (strcmp(s, "stddev") == 0)   st->want_stddev = 1;
+            else if (strcmp(s, "median") == 0)   st->want_median = 1;
+            else if (strcmp(s, "p25") == 0)      st->want_p25 = 1;
+            else if (strcmp(s, "p75") == 0)      st->want_p75 = 1;
+            else if (strcmp(s, "skewness") == 0) st->want_skewness = 1;
+            else if (strcmp(s, "kurtosis") == 0) st->want_kurtosis = 1;
+            else if (strcmp(s, "distinct") == 0) st->want_distinct = 1;
+            else if (strcmp(s, "hist") == 0)     st->want_hist = 1;
+            else if (strcmp(s, "sample") == 0)   st->want_sample = 1;
+        }
+    } else {
+        /* Default: basic + variance + median */
+        st->want_count = 1;
+        st->want_sum = 1;
+        st->want_avg = 1;
+        st->want_min = 1;
+        st->want_max = 1;
+        st->want_var = 1;
+        st->want_stddev = 1;
+        st->want_median = 1;
+    }
+
+    tf_step *step = malloc(sizeof(tf_step));
+    if (!step) { free(st); return NULL; }
+    step->process = stats_process;
+    step->flush = stats_flush;
+    step->destroy = stats_destroy;
+    step->state = st;
+    return step;
+}