Near real-time index updates

While index updates are fast on average, worst-case update performance may be significantly slower. To make index updates consistently quick, we must manually intervene to control the process of index segment consolidation.

The problem

Ordinarily, modifying an index is cheap. New data is added to new segments, and the time to write a new segment scales more or less linearly with the number of documents added during the indexing session.

Deletions are also cheap most of the time, because we don’t remove documents immediately but instead mark them as deleted, and adding the deletion mark is cheap.

However, as new segments are added and the deletion rate for existing segments increases, search-time performance slowly begins to degrade. At some point, it becomes necessary to consolidate existing segments, rewriting their data into a new segment.

If the recycled segments are small, the time it takes to rewrite them may not be significant. Every once in a while, though, a large amount of data must be rewritten.

Procrastinating and playing catch-up

The simplest way to force fast index updates is to avoid rewriting anything.

Indexer relies upon IndexManager’s Recycle() method to tell it which segments should be consolidated. If we subclass IndexManager and override the method so that it always returns an empty array, we get consistently quick performance:

Vector*
NoMergeManager_Recycle_IMP(IndexManager *self, PolyReader *reader,
                           DeletionsWriter *del_writer, int64_t cutoff,
                           bool optimize) {
    return Vec_new(0);
}

void
do_index(Obj *index) {
    CFCClass *klass = Class_singleton("NoMergeManager", INDEXMANAGER);
    Class_Override(klass, (cfish_method_t)NoMergeManager_Recycle_IMP,
                   LUCY_IndexManager_Recycle_OFFSET);

    IndexManager *manager = (IndexManager*)Class_Make_Obj(klass);
    IxManager_init(manager, NULL, NULL);

    Indexer *indexer = Indexer_new(NULL, index, manager, 0);
    ...
    Indexer_Commit(indexer);

    DECREF(indexer);
    DECREF(manager);
}

However, we can’t procrastinate forever. Eventually, we’ll have to run an ordinary, uncontrolled indexing session, potentially triggering a large rewrite of lots of small and/or degraded segments:

void
do_index(Obj *index) {
    Indexer *indexer = Indexer_new(NULL, index, NULL /* manager */, 0);
    ...
    Indexer_Commit(indexer);
    DECREF(indexer);
}

Acceptable worst-case update time, slower degradation

Never merging anything at all in the main indexing process is probably overkill. Small segments are relatively cheap to merge; we just need to guard against the big rewrites.

Setting a ceiling on the number of documents in the segments to be recycled allows us to avoid a mass proliferation of tiny, single-document segments, while still offering decent worst-case update speed:

Vector*
LightMergeManager_Recycle_IMP(IndexManager *self, PolyReader *reader,
                              DeletionsWriter *del_writer, int64_t cutoff,
                              bool optimize) {
    IndexManager_Recycle_t super_recycle
        = SUPER_METHOD_PTR(IndexManager, LUCY_IndexManager_Recycle);
    Vector *seg_readers = super_recycle(self, reader, del_writer, cutoff,
                                        optimize);
    Vector *small_segments = Vec_new(0);

    for (size_t i = 0, max = Vec_Get_Size(seg_readers); i < max; i++) {
        SegReader *seg_reader = (SegReader*)Vec_Fetch(seg_readers, i);

        if (SegReader_Doc_Max(seg_reader) < 10) {
            Vec_Push(small_segments, INCREF(seg_reader));
        }
    }

    DECREF(seg_readers);
    return small_segments;
}

However, we still have to consolidate every once in a while, and while that happens content updates will be locked out.

Background merging

If it’s not acceptable to lock out updates while the index consolidation process runs, the alternative is to move the consolidation process out of band, using BackgroundMerger.

It’s never safe to have more than one Indexer attempting to modify the content of an index at the same time, but a BackgroundMerger and an Indexer can operate simultaneously:

typedef struct {
    Obj *index;
    Doc *doc;
} Context;

static void
S_index_doc(void *arg) {
    Context *ctx = (Context*)arg;

    CFCClass *klass = Class_singleton("LightMergeManager", INDEXMANAGER);
    Class_Override(klass, (cfish_method_t)LightMergeManager_Recycle_IMP,
                   LUCY_IndexManager_Recycle_OFFSET);

    IndexManager *manager = (IndexManager*)Class_Make_Obj(klass);
    IxManager_init(manager, NULL, NULL);

    Indexer *indexer = Indexer_new(NULL, ctx->index, manager, 0);
    Indexer_Add_Doc(indexer, ctx->doc, 1.0);
    Indexer_Commit(indexer);

    DECREF(indexer);
    DECREF(manager);
}

void indexing_process(Obj *index, Doc *doc) {
    Context ctx;
    ctx.index = index;
    ctx.doc = doc;

    for (int i = 0; i < max_retries; i++) {
        Err *err = Err_trap(S_index_doc, &ctx);
        if (!err) { break; }
        if (!Err_is_a(err, LOCKERR)) {
            RETHROW(err);
        }
        WARN("Couldn't get lock (%d retries)", i);
        DECREF(err);
    }
}

void
background_merge_process(Obj *index) {
    IndexManager *manager = IxManager_new(NULL, NULL);
    IxManager_Set_Write_Lock_Timeout(manager, 60000);

    BackgroundMerger bg_merger = BGMerger_new(index, manager);
    BGMerger_Commit(bg_merger);

    DECREF(bg_merger);
    DECREF(manager);
}

The exception handling code becomes useful once you have more than one index modification process happening simultaneously. By default, Indexer tries several times to acquire a write lock over the span of one second, then holds it until Commit() completes. BackgroundMerger handles most of its work without the write lock, but it does need it briefly once at the beginning and once again near the end. Under normal loads, the internal retry logic will resolve conflicts, but if it’s not acceptable to miss an insert, you probably want to catch LockErr exceptions thrown by Indexer. In contrast, a LockErr from BackgroundMerger probably just needs to be logged.