Overview
- Major Objects
- Use Major Objects for fast in-memory handling of large amount of data that is thread-safe but must be persisted
- We must support complex objects with simple keys, crud, and fast lookup by multiple keys.
- Use an unordered_set of const pointers to objects derived from PersistentIDObject (see below)
- The main container's primary key is always db_id
- Always use the db_id for foreign keys
- Other containers can be created using other members as keys; the only cost is for a new set of pointers (not objects!)
- For other unordered_sets, just define new hash functions
- Other useful containers are sorted_vector and map (<key,value> pair)
- PersistentIDObject
- Add a dirty flag to all objects, set to true on any change that must be persisted
- Use an internal in-memory counter to generate the next db_id for a newly created object
- This means that when creating new objects, there is NO NEED to access db, VERY IMPORTANT!
- Use delayed-write tactics to write all dirty objects on idle time
- Memory Model
- Use a Datastore manager (aka "MemoryModel") to hold sets
- It can look up objects by any key, and strip away const to return a mutable object. NOTE that the user must not damage the key values!
- Derive a class from the memory model for persistence; it can use any persistence method (local, remote, sql, nosql, etc.).
- Make sure that the base MemoryModel class is concrete not abstract, thread-safe and self-contained; this makes parallel calculations trivial, helps scalability, etc.
- Schema Driven
- Schema is shared across memory objects and every persistence layer. Follow a DRY pattern by generating all code from a common schema definition.
- JSON schema is elegant as it is as simple as possible but no less, always use it. Essential tools:
- Use unsorted_set<PersistentIDObject*> containers that can be automatically serialized to/from any database layer
- Always provide a constructor with a single db_id parameter, defaulted to UNSAVED; this serves two purposes:
- no-param constructor for reflection via quicktype
- id constructor for use as key for unsorted_set::find()
Delayed delete pattern
1) to dynamically delete an object:
a) ba.setDeleted();
b) do not remove from any container indexes
c) but fix the index sorting, flags, etc, as if the object were gone, so the program will function as if it is!
eg: do not remove from runsByRank_, but adjust all other ranking as if the run was gone
2) include deleted status in active check, etc.:
// NOTE use the direct function rather than !bFunc(), as deleted objects return false for both.
bool bActive() const { return b_active_ && !bDeleted(); }
bool bInactive() const { return !b_active_ && !bDeleted(); }
---
for (auto& psr: runsByRank_) {
if (psr->bDeleted()) continue;
...
3) all deletion work is done in MemoryModel::saveDirtyObjectsAsNeeded(), see that code
a) deletion check should happen in delayed write check:
if (pau->bDirtyOrDeleted())
bNeeded = true;
b) if bNeeded, always do deletions first, starting with greatest grandparent container, to minimize work
c) use the erase-remove pattern to remove all deleted items in one loop
see code here for reference implementation: BrokerAccount::removeDeletedStockRuns()
i) iterate and remove item from all secondary indices
ii) iterate primary index, and use the lambda of the erase-remove operation to delete memory allocation and remove db record
iii) associative container iterators can be safely deleted directly
sequential containers like vector require use of erase-remove idiom
see reference implementation for example code!