diff --git a/.golangci.yml b/.golangci.yml
index 78effdf9..5f112400 100644
--- a/.golangci.yml
+++ b/.golangci.yml
@@ -16,12 +16,18 @@ linters-settings:
     values:
       regexp:
         copyright-year: 20[2-9]\d
-
+  forbidigo:
+    forbid:
+      - p: ^pgxpool\.New.*$
+        msg: "Use github.com/conduitio/conduit-connector-postgres/source/cpool.New instead."
 issues:
   exclude-rules:
     - path: test/helper\.go
       linters:
         - gosec
+    - path: source/cpool/cpool\.go
+      linters:
+      - forbidigo
 
 linters:
   # please, do not use `enable-all`: it's deprecated and will be removed soon.
@@ -38,7 +44,7 @@ linters:
     # - exhaustive
     # - exhaustivestruct
     - exportloopref
-    # - forbidigo
+    - forbidigo
     # - forcetypeassert
     # - funlen
     # - gochecknoinits
diff --git a/source/cpool/cpool.go b/source/cpool/cpool.go
index 224f4df3..ed077411 100644
--- a/source/cpool/cpool.go
+++ b/source/cpool/cpool.go
@@ -59,6 +59,10 @@ func beforeAcquireHook(ctx context.Context, conn *pgx.Conn) bool {
 
 // beforeConnectHook customizes the configuration of the new connection.
 func beforeConnectHook(ctx context.Context, config *pgx.ConnConfig) error {
+	if config.RuntimeParams["application_name"] == "" {
+		config.RuntimeParams["application_name"] = "conduit-connector-postgres"
+	}
+
 	if v := ctx.Value(replicationCtxKey{}); v != nil {
 		config.RuntimeParams["replication"] = "database"
 	}
diff --git a/source/logrepl/cleaner.go b/source/logrepl/cleaner.go
index befa20fa..db52c9fc 100644
--- a/source/logrepl/cleaner.go
+++ b/source/logrepl/cleaner.go
@@ -19,9 +19,9 @@ import (
 	"errors"
 	"fmt"
 
+	"github.com/conduitio/conduit-connector-postgres/source/cpool"
 	"github.com/conduitio/conduit-connector-postgres/source/logrepl/internal"
 	sdk "github.com/conduitio/conduit-connector-sdk"
-	"github.com/jackc/pgx/v5/pgxpool"
 )
 
 type CleanupConfig struct {
@@ -35,7 +35,7 @@ type CleanupConfig struct {
 func Cleanup(ctx context.Context, c CleanupConfig) error {
 	logger := sdk.Logger(ctx)
 
-	pool, err := pgxpool.New(ctx, c.URL)
+	pool, err := cpool.New(ctx, c.URL)
 	if err != nil {
 		return fmt.Errorf("failed to connect to database: %w", err)
 	}
diff --git a/source/logrepl/combined_test.go b/source/logrepl/combined_test.go
index 592efebf..4abca0da 100644
--- a/source/logrepl/combined_test.go
+++ b/source/logrepl/combined_test.go
@@ -165,6 +165,16 @@ func TestCombinedIterator_Next(t *testing.T) {
 
 	expectedRecords := testRecords()
 
+	// interrupt repl connection
+	var terminated bool
+	is.NoErr(pool.QueryRow(ctx, fmt.Sprintf(
+		`SELECT pg_terminate_backend(pid) FROM pg_stat_activity WHERE
+			query ILIKE '%%CREATE_REPLICATION_SLOT %s%%' and pid <> pg_backend_pid()
+		`,
+		table,
+	)).Scan(&terminated))
+	is.True(terminated)
+
 	// compare snapshot
 	for id := 1; id < 5; id++ {
 		t.Run(fmt.Sprint("next_snapshot", id), func(t *testing.T) {
diff --git a/source/logrepl/internal/subscription.go b/source/logrepl/internal/subscription.go
index a7e0d0da..193356ce 100644
--- a/source/logrepl/internal/subscription.go
+++ b/source/logrepl/internal/subscription.go
@@ -44,6 +44,7 @@ type Subscription struct {
 	TXSnapshotID  string
 
 	conn *pgxpool.Conn
+	pool *pgxpool.Pool
 
 	stop context.CancelFunc
 
@@ -133,6 +134,7 @@ func CreateSubscription(
 		TXSnapshotID:  result.SnapshotName,
 
 		conn: conn,
+		pool: pool,
 
 		ready: make(chan struct{}),
 		done:  make(chan struct{}),
@@ -144,6 +146,7 @@ func (s *Subscription) Run(ctx context.Context) error {
 	defer s.doneReplication()
 
 	if err := s.startReplication(ctx); err != nil {
+		close(s.ready) // ready to fail.
 		return err
 	}
 
@@ -330,6 +333,18 @@ func (s *Subscription) Err() error {
 
 // startReplication starts replication with a specific start LSN.
 func (s *Subscription) startReplication(ctx context.Context) error {
+	// N.B. Snapshots may take long time and connection may timeout.
+	// 		Safer to refresh the connection before replication begins.
+
+	s.conn.Release()
+
+	conn, err := s.pool.Acquire(cpool.WithReplication(ctx))
+	if err != nil {
+		return fmt.Errorf("could not establish replication connection: %w", err)
+	}
+
+	s.conn = conn
+
 	pluginArgs := []string{
 		`"proto_version" '1'`,
 		fmt.Sprintf(`"publication_names" '%s'`, s.Publication),
diff --git a/source/schema/avro.go b/source/schema/avro.go
index ac119d0f..112b853b 100644
--- a/source/schema/avro.go
+++ b/source/schema/avro.go
@@ -17,7 +17,6 @@ package schema
 import (
 	"cmp"
 	"fmt"
-	"math"
 	"slices"
 
 	"github.com/hamba/avro/v2"
@@ -27,15 +26,8 @@ import (
 )
 
 const (
-	avroNS = "conduit.postgres"
-	// The default decimal precision is pretty generous, but it is in excess of what
-	// pgx provides by default. All numeric values by default are coded to float64/int64.
-	// Ideally in the future the decimal precision can be adjusted to fit the definition in postgres.
-	avroDecimalPrecision = 38
-	// The size of the storage in which a decimal may be encoded depends on the underlying numeric definition.
-	// Unfortunately similarly to the decimal precision, this is dependent on the size of the numeric, which
-	// by default is constraint to 8 bytes. This default is generously allocating four times larger width.
-	avroDecimalFixedSize = 8 * 4
+	avroNS             = "conduit.postgres"
+	avroDecimalPadding = 8
 )
 
 var Avro = &avroExtractor{
@@ -75,38 +67,29 @@ type avroExtractor struct {
 }
 
 func (a avroExtractor) ExtractLogrepl(rel *pglogrepl.RelationMessage, row *pglogrepl.TupleData) (avro.Schema, error) {
-	var (
-		fields []pgconn.FieldDescription
-		values []any
-	)
+	var fields []pgconn.FieldDescription
 
-	for i, tuple := range row.Columns {
+	for i := range row.Columns {
 		fields = append(fields, pgconn.FieldDescription{
-			Name:        rel.Columns[i].Name,
-			DataTypeOID: rel.Columns[i].DataType,
+			Name:         rel.Columns[i].Name,
+			DataTypeOID:  rel.Columns[i].DataType,
+			TypeModifier: rel.Columns[i].TypeModifier,
 		})
-
-		v, err := a.decodeColumnValue(rel.Columns[i], tuple.Data)
-		if err != nil {
-			return nil, err
-		}
-
-		values = append(values, v)
 	}
 
-	return a.Extract(rel.RelationName, fields, values)
+	return a.Extract(rel.RelationName, fields)
 }
 
-func (a *avroExtractor) Extract(name string, fields []pgconn.FieldDescription, values []any) (avro.Schema, error) {
+func (a *avroExtractor) Extract(name string, fields []pgconn.FieldDescription) (avro.Schema, error) {
 	var avroFields []*avro.Field
 
-	for i, f := range fields {
+	for _, f := range fields {
 		t, ok := a.pgMap.TypeForOID(f.DataTypeOID)
 		if !ok {
 			return nil, fmt.Errorf("field %q with OID %d cannot be resolved", f.Name, f.DataTypeOID)
 		}
 
-		s, err := a.extractType(t, values[i])
+		s, err := a.extractType(t, f.TypeModifier)
 		if err != nil {
 			return nil, err
 		}
@@ -131,41 +114,26 @@ func (a *avroExtractor) Extract(name string, fields []pgconn.FieldDescription, v
 	return sch, nil
 }
 
-func (a *avroExtractor) extractType(t *pgtype.Type, val any) (avro.Schema, error) {
+func (a *avroExtractor) extractType(t *pgtype.Type, typeMod int32) (avro.Schema, error) {
 	if ps, ok := a.avroMap[t.Name]; ok {
 		return ps, nil
 	}
 
-	switch tt := val.(type) {
-	case pgtype.Numeric:
-		// N.B.: Default to 38 positions and pick the exponent as the scale.
+	switch t.OID {
+	case pgtype.NumericOID:
+		scale := int((typeMod - 4) & 65535)
+		precision := int(((typeMod - 4) >> 16) & 65535)
 		fs, err := avro.NewFixedSchema(
 			string(avro.Decimal),
 			avroNS,
-			avroDecimalFixedSize,
-			avro.NewDecimalLogicalSchema(avroDecimalPrecision, int(math.Abs(float64(tt.Exp)))),
+			precision+scale+avroDecimalPadding,
+			avro.NewDecimalLogicalSchema(precision, scale),
 		)
 		if err != nil {
 			return nil, fmt.Errorf("failed to create avro.FixedSchema: %w", err)
 		}
 		return fs, nil
 	default:
-		return nil, fmt.Errorf("cannot resolve field %q of type %T", t.Name, tt)
+		return nil, fmt.Errorf("cannot resolve field type %q ", t.Name)
 	}
 }
-
-func (a *avroExtractor) decodeColumnValue(col *pglogrepl.RelationMessageColumn, data []byte) (any, error) {
-	var t *pgtype.Type
-
-	t, ok := a.pgMap.TypeForOID(col.DataType)
-	if !ok {
-		t, _ = a.pgMap.TypeForOID(pgtype.UnknownOID)
-	}
-
-	v, err := t.Codec.DecodeValue(a.pgMap, col.DataType, pgtype.TextFormatCode, data)
-	if err != nil {
-		return nil, fmt.Errorf("failed to decode %q tuple: %w", col.Name, err)
-	}
-
-	return v, nil
-}
diff --git a/source/schema/avro_test.go b/source/schema/avro_test.go
index 679597e0..c5d2176c 100644
--- a/source/schema/avro_test.go
+++ b/source/schema/avro_test.go
@@ -50,7 +50,7 @@ func Test_AvroExtract(t *testing.T) {
 
 	fields := rows.FieldDescriptions()
 
-	sch, err := Avro.Extract(table, fields, values)
+	sch, err := Avro.Extract(table, fields)
 
 	t.Run("schema is parsable", func(t *testing.T) {
 		is := is.New(t)
@@ -191,8 +191,8 @@ func avroTestSchema(t *testing.T, table string) avro.Schema {
 		assert(avro.NewField("col_numeric",
 			assert(avro.NewFixedSchema(string(avro.Decimal),
 				avroNS,
-				avroDecimalFixedSize,
-				avro.NewDecimalLogicalSchema(avroDecimalPrecision, 2),
+				18,
+				avro.NewDecimalLogicalSchema(8, 2),
 			)))),
 		assert(avro.NewField("col_date", avro.NewPrimitiveSchema(
 			avro.Int,
diff --git a/source/snapshot/fetch_worker.go b/source/snapshot/fetch_worker.go
index 6a64a72b..b57e9696 100644
--- a/source/snapshot/fetch_worker.go
+++ b/source/snapshot/fetch_worker.go
@@ -286,7 +286,7 @@ func (f *FetchWorker) fetch(ctx context.Context, tx pgx.Tx) (int, error) {
 		}
 
 		if f.conf.WithAvroSchema && f.avroSchema == nil {
-			sch, err := schema.Avro.Extract(f.conf.Table, fields, values)
+			sch, err := schema.Avro.Extract(f.conf.Table, fields)
 			if err != nil {
 				return 0, fmt.Errorf("failed to extract schema: %w", err)
 			}