Tarantool filer store (#6669)

Co-authored-by: Marat Karimov <m.karimov@digitalms.ru>

8 files changed, 395 insertions, 2 deletions

diff --git a/docker/Dockerfile.tarantool.dev_env b/docker/Dockerfile.tarantool.dev_env
new file mode 100644
index 000000000..4ce0fc9af
--- /dev/null
+++ b/docker/Dockerfile.tarantool.dev_env
@@ -0,0 +1,17 @@
+FROM tarantool/tarantool:3.3.1 AS builder
+
+# install dependencies
+RUN apt update && \
+  apt install -y git unzip cmake tt=2.7.0
+
+# init tt dir structure, create dir for app, create symlink
+RUN tt init && \
+  mkdir app && \
+  ln -sfn ${PWD}/app/ ${PWD}/instances.enabled/app
+
+# copy cluster configs
+COPY tarantool /opt/tarantool/app
+
+# build app
+RUN tt build app
+
diff --git a/docker/Makefile b/docker/Makefile
index d4dd70b6b..a4f207c89 100644
--- a/docker/Makefile
+++ b/docker/Makefile
@@ -22,7 +22,7 @@ build: binary
 build_e2e: binary_race
 	docker build --no-cache -t chrislusf/seaweedfs:e2e -f Dockerfile.e2e .
 
-go_build: # make go_build tags=elastic,ydb,gocdk,hdfs,5BytesOffset
+go_build: # make go_build tags=elastic,ydb,gocdk,hdfs,5BytesOffset,tarantool
 	docker build --build-arg TAGS=$(tags) --no-cache -t chrislusf/seaweedfs:go_build -f Dockerfile.go_build .
 
 go_build_large_disk:
@@ -37,6 +37,9 @@ build_rocksdb_local: build_rocksdb_dev_env
 build_rocksdb:
 	docker build --no-cache -t chrislusf/seaweedfs:rocksdb -f Dockerfile.rocksdb_large .
 
+build_tarantool_dev_env:
+	docker build --no-cache -t chrislusf/tarantool_dev_env -f Dockerfile.tarantool.dev_env .
+
 s3tests_build:
 	docker build --no-cache -t chrislusf/ceph-s3-tests:local -f Dockerfile.s3tests .
 
@@ -106,9 +109,12 @@ test_etcd: build
 
 test_ydb: tags = ydb
 test_ydb: build
-	export
 	docker compose -f compose/test-ydb-filer.yml -p seaweedfs up
 
+test_tarantool: tags = tarantool
+test_tarantool: build_tarantool_dev_env build
+	docker compose -f compose/test-tarantool-filer.yml -p seaweedfs up
+
 clean:
 	rm ./weed
 
diff --git a/docker/compose/test-tarantool-filer.yml b/docker/compose/test-tarantool-filer.yml
new file mode 100644
index 000000000..8f31bf855
--- /dev/null
+++ b/docker/compose/test-tarantool-filer.yml
@@ -0,0 +1,30 @@
+version: '3.9'
+
+services:
+  tarantool:
+    image: chrislusf/tarantool_dev_env
+    entrypoint: "tt start app -i"
+    environment:
+      APP_USER_PASSWORD: "app"
+      CLIENT_USER_PASSWORD: "client"
+      REPLICATOR_USER_PASSWORD: "replicator"
+      STORAGE_USER_PASSWORD: "storage"
+    network_mode: "host"
+    ports:
+      - "3303:3303"
+
+  s3:
+    image: chrislusf/seaweedfs:local
+    command: "server -ip=127.0.0.1 -filer -master.volumeSizeLimitMB=16 -volume.max=0 -volume -volume.preStopSeconds=1 -s3 -s3.config=/etc/seaweedfs/s3.json -s3.port=8000 -s3.allowEmptyFolder=false -s3.allowDeleteBucketNotEmpty=false"
+    volumes:
+      - ./s3.json:/etc/seaweedfs/s3.json
+    environment:
+      WEED_LEVELDB2_ENABLED: "false"
+      WEED_TARANTOOL_ENABLED: "true"
+      WEED_TARANTOOL_ADDRESS: "127.0.0.1:3303"
+      WEED_TARANTOOL_USER: "client"
+      WEED_TARANTOOL_PASSWORD: "client"
+      WEED_MASTER_VOLUME_GROWTH_COPY_OTHER: 1
+    network_mode: "host"
+    depends_on:
+      - tarantool
\ No newline at end of file
diff --git a/docker/tarantool/app-scm-1.rockspec b/docker/tarantool/app-scm-1.rockspec
new file mode 100644
index 000000000..79eb1ca38
--- /dev/null
+++ b/docker/tarantool/app-scm-1.rockspec
@@ -0,0 +1,14 @@
+package = 'app'
+version = 'scm-1'
+source  = {
+    url = '/dev/null',
+}
+dependencies = {
+    'crud == 1.5.2-1',
+    'expirationd == 1.6.0-1',
+    'metrics-export-role == 0.3.0-1',
+    'vshard == 0.1.32-1'
+}
+build = {
+    type = 'none';
+}
\ No newline at end of file
diff --git a/docker/tarantool/config.yaml b/docker/tarantool/config.yaml
new file mode 100644
index 000000000..00a693a2e
--- /dev/null
+++ b/docker/tarantool/config.yaml
@@ -0,0 +1,145 @@
+config:
+  context:
+    app_user_password:
+      from: env
+      env: APP_USER_PASSWORD
+    client_user_password:
+      from: env
+      env: CLIENT_USER_PASSWORD
+    replicator_user_password:
+      from: env
+      env: REPLICATOR_USER_PASSWORD
+    storage_user_password:
+      from: env
+      env: STORAGE_USER_PASSWORD
+
+credentials:
+  roles:
+    crud-role:
+      privileges:
+        - permissions: [ "execute" ]
+          lua_call: [ "crud.delete", "crud.get", "crud.upsert" ]
+  users:
+    app:
+      password: '{{ context.app_user_password }}'
+      roles: [ public, crud-role ]
+    client:
+      password: '{{ context.client_user_password }}'
+      roles: [ super ]
+    replicator:
+      password: '{{ context.replicator_user_password }}'
+      roles: [ replication ]
+    storage:
+      password: '{{ context.storage_user_password }}'
+      roles: [ sharding ]
+
+iproto:
+  advertise:
+    peer:
+      login: replicator
+    sharding:
+      login: storage
+
+sharding:
+  bucket_count: 10000
+
+metrics:
+  include: [ all ]
+  exclude: [ vinyl ]
+  labels:
+    alias: '{{ instance_name }}'
+
+
+groups:
+  storages:
+    roles:
+      - roles.crud-storage
+      - roles.expirationd
+      - roles.metrics-export
+    roles_cfg:
+      roles.expirationd:
+        cfg:
+          metrics: true
+        filer_metadata_task:
+          space: filer_metadata
+          is_expired: filer_metadata.is_expired
+          options:
+            atomic_iteration: true
+            force: true
+            index: 'expire_at_idx'
+            iterator_type: GT
+            start_key:
+              - 0
+            tuples_per_iteration: 10000
+    app:
+      module: storage
+    sharding:
+      roles: [ storage ]
+    replication:
+      failover: election
+    database:
+      use_mvcc_engine: true
+    replicasets:
+      storage-001:
+        instances:
+          storage-001-a:
+            roles_cfg:
+              roles.metrics-export:
+                http:
+                  - listen: '0.0.0.0:8081'
+                    endpoints:
+                      - path: /metrics/prometheus/
+                        format: prometheus
+                      - path: /metrics/json
+                        format: json
+            iproto:
+              listen:
+                - uri: 127.0.0.1:3301
+              advertise:
+                client: 127.0.0.1:3301
+          storage-001-b:
+            roles_cfg:
+              roles.metrics-export:
+                http:
+                  - listen: '0.0.0.0:8082'
+                    endpoints:
+                      - path: /metrics/prometheus/
+                        format: prometheus
+                      - path: /metrics/json
+                        format: json
+            iproto:
+              listen:
+                - uri: 127.0.0.1:3302
+              advertise:
+                client: 127.0.0.1:3302
+  routers:
+    roles:
+      - roles.crud-router
+      - roles.metrics-export
+    roles_cfg:
+      roles.crud-router:
+        stats: true
+        stats_driver: metrics
+        stats_quantiles: true
+    app:
+      module: router
+    sharding:
+      roles: [ router ]
+    replicasets:
+      router-001:
+        instances:
+          router-001-a:
+            roles_cfg:
+              roles.metrics-export:
+                http:
+                  - listen: '0.0.0.0:8083'
+                    endpoints:
+                      - path: /metrics/prometheus/
+                        format: prometheus
+                      - path: /metrics/json
+                        format: json
+            iproto:
+              listen:
+                - uri: 127.0.0.1:3303
+              advertise:
+                client: 127.0.0.1:3303
\ No newline at end of file
diff --git a/docker/tarantool/instances.yaml b/docker/tarantool/instances.yaml
new file mode 100644
index 000000000..225b7382f
--- /dev/null
+++ b/docker/tarantool/instances.yaml
@@ -0,0 +1,7 @@
+---
+storage-001-a:
+
+storage-001-b:
+
+router-001-a:
+
diff --git a/docker/tarantool/router.lua b/docker/tarantool/router.lua
new file mode 100644
index 000000000..359a8c49b
--- /dev/null
+++ b/docker/tarantool/router.lua
@@ -0,0 +1,77 @@
+local vshard = require('vshard')
+local log = require('log')
+
+-- Bootstrap the vshard router.
+while true do
+    local ok, err = vshard.router.bootstrap({
+        if_not_bootstrapped = true,
+    })
+    if ok then
+        break
+    end
+    log.info(('Router bootstrap error: %s'):format(err))
+end
+
+-- functions for filer_metadata space
+local filer_metadata = {
+    delete_by_directory_idx = function(directory)
+        -- find all storages
+        local storages = require('vshard').router.routeall()
+        -- on each storage
+        for _, storage in pairs(storages) do
+            -- call local function
+            local result, err = storage:callrw('filer_metadata.delete_by_directory_idx', { directory })
+            -- check for error
+            if err then
+                error("Failed to call function on storage: " .. tostring(err))
+            end
+        end
+        -- return
+        return true
+    end,
+    find_by_directory_idx_and_name = function(dirPath, startFileName, includeStartFile, limit)
+        -- init results
+        local results = {}
+        -- find all storages
+        local storages = require('vshard').router.routeall()
+        -- on each storage
+        for _, storage in pairs(storages) do
+            -- call local function
+            local result, err = storage:callro('filer_metadata.find_by_directory_idx_and_name', {
+                dirPath,
+                startFileName,
+                includeStartFile,
+                limit
+            })
+            -- check for error
+            if err then
+                error("Failed to call function on storage: " .. tostring(err))
+            end
+            -- add to results
+            for _, tuple in ipairs(result) do
+                table.insert(results, tuple)
+            end
+        end
+        -- sort
+        table.sort(results, function(a, b) return a[3] < b[3] end)
+        -- apply limit
+        if #results > limit then
+            local limitedResults = {}
+            for i = 1, limit do
+                table.insert(limitedResults, results[i])
+            end
+            results = limitedResults
+        end
+        -- return
+        return results
+    end,
+}
+
+rawset(_G, 'filer_metadata', filer_metadata)
+
+-- register functions for filer_metadata space, set grants
+for name, _ in pairs(filer_metadata) do
+    box.schema.func.create('filer_metadata.' .. name, { if_not_exists = true })
+    box.schema.user.grant('app', 'execute', 'function', 'filer_metadata.' .. name, { if_not_exists = true })
+    box.schema.user.grant('client', 'execute', 'function', 'filer_metadata.' .. name, { if_not_exists = true })
+end
diff --git a/docker/tarantool/storage.lua b/docker/tarantool/storage.lua
new file mode 100644
index 000000000..ff1ec0288
--- /dev/null
+++ b/docker/tarantool/storage.lua
@@ -0,0 +1,97 @@
+box.watch('box.status', function()
+    if box.info.ro then
+        return
+    end
+
+    -- ====================================
+    -- key_value space
+    -- ====================================
+    box.schema.create_space('key_value', {
+        format = {
+            { name = 'key', type = 'string' },
+            { name = 'bucket_id', type = 'unsigned' },
+            { name = 'value', type = 'string' }
+        },
+        if_not_exists = true
+    })
+
+    -- create key_value space indexes
+    box.space.key_value:create_index('id', {type = 'tree', parts = { 'key' }, unique = true, if_not_exists = true})
+    box.space.key_value:create_index('bucket_id', { type = 'tree', parts = { 'bucket_id' }, unique = false, if_not_exists = true })
+
+    -- ====================================
+    -- filer_metadata space
+    -- ====================================
+    box.schema.create_space('filer_metadata', {
+        format = {
+            { name = 'directory', type = 'string' },
+            { name = 'bucket_id', type = 'unsigned' },
+            { name = 'name', type = 'string' },
+            { name = 'expire_at', type = 'unsigned' },
+            { name = 'data', type = 'string' }
+        },
+        if_not_exists = true
+    })
+
+    -- create filer_metadata space indexes
+    box.space.filer_metadata:create_index('id', {type = 'tree', parts = { 'directory', 'name' }, unique = true, if_not_exists = true})
+    box.space.filer_metadata:create_index('bucket_id', { type = 'tree', parts = { 'bucket_id' }, unique = false, if_not_exists = true })
+    box.space.filer_metadata:create_index('directory_idx', { type = 'tree', parts = { 'directory' }, unique = false, if_not_exists = true })
+    box.space.filer_metadata:create_index('name_idx', { type = 'tree', parts = { 'name' }, unique = false, if_not_exists = true })
+    box.space.filer_metadata:create_index('expire_at_idx', { type = 'tree', parts = { 'expire_at' }, unique = false, if_not_exists = true})
+end)
+
+-- functions for filer_metadata space
+local filer_metadata = {
+    delete_by_directory_idx = function(directory)
+        local space = box.space.filer_metadata
+        local index = space.index.directory_idx
+        -- for each finded directories
+        for _, tuple in index:pairs({ directory }, { iterator = 'EQ' }) do
+            space:delete({ tuple[1], tuple[3] })
+        end
+        return true
+    end,
+    find_by_directory_idx_and_name = function(dirPath, startFileName, includeStartFile, limit)
+        local space = box.space.filer_metadata
+        local directory_idx = space.index.directory_idx
+        -- choose filter name function
+        local filter_filename_func
+        if includeStartFile then
+            filter_filename_func = function(value) return value >= startFileName end
+        else
+            filter_filename_func = function(value) return value > startFileName end
+        end
+        -- init results
+        local results = {}
+        -- for each finded directories
+        for _, tuple in directory_idx:pairs({ dirPath }, { iterator = 'EQ' }) do
+            -- filter by name
+            if filter_filename_func(tuple[3]) then
+                table.insert(results, tuple)
+            end
+        end
+        -- sort
+        table.sort(results, function(a, b) return a[3] < b[3] end)
+        -- apply limit
+        if #results > limit then
+            local limitedResults = {}
+            for i = 1, limit do
+                table.insert(limitedResults, results[i])
+            end
+            results = limitedResults
+        end
+        -- return
+        return results
+    end,
+    is_expired = function(args, tuple)
+        return (tuple[4] > 0) and (require('fiber').time() > tuple[4])
+    end
+}
+
+-- register functions for filer_metadata space, set grants
+rawset(_G, 'filer_metadata', filer_metadata)
+for name, _ in pairs(filer_metadata) do
+    box.schema.func.create('filer_metadata.' .. name, { setuid = true, if_not_exists = true })
+    box.schema.user.grant('storage', 'execute', 'function', 'filer_metadata.' .. name, { if_not_exists = true })
+end